Changelog

New white papers and new LLM, VLM and CV benchmark results, newest first. Each entry links to what it describes.

November 2026

  1. 17 NovemberNew VLM benchmarkAlpamayo2-Super — Jetson AGX ThorAlpamayo2-Super on Jetson AGX Thor: the vision-language part in BF16 served with vLLM, response time measured from 1 to 176 simultaneous cameras and up to 256 images per request, with a six-camera driving test of the whole model using its official PyTorch inference (13.3 s per decision).
  2. 17 NovemberNew VLM benchmarkAlpamayo2-Edge — Jetson AGX ThorAlpamayo2-Edge on Jetson AGX Thor: the vision-language part in BF16 served with vLLM, response time measured from 1 to 1,408 simultaneous cameras and up to 512 images per request, with a six-camera driving test of the whole model using its TensorRT Edge-LLM recipe (331 ms per decision).
  3. 13 NovemberNew VLM benchmarkQwen3-VL-30B-A3B-Instruct — DGX B300 (1 GPU)Qwen3-VL-30B-A3B-Instruct on DGX B300 (1 GPU): FP8 weights served with vLLM, response time measured from 1 to 384 simultaneous cameras and up to 256 images per request.
  4. 9 NovemberNew LLM benchmarkKimi-K3.1-Pro — DGX B300Kimi-K3.1-Pro on DGX B300: MXFP4 weights served with SGLang across 8 GPUs, with DFlash2 speculative decoding (k=8) and a 1M context, measured up to 128 concurrent requests.
  5. 6 NovemberNew LLM benchmarkMistral-Large-4-1T-A52B — DGX B300Mistral-Large-4-1T-A52B on DGX B300: FP8 weights served with SGLang across 8 GPUs, with DSpark speculative decoding (k=3) and a 1M context, measured up to 128 concurrent requests.
  6. 6 NovemberNew LLM benchmarkRipple-65B-Preview — DGX B300Ripple-65B-Preview on DGX B300: NVFP4 weights served with SGLang on one GPU, with DFlash2 speculative decoding (k=16) and a 2M context, measured up to 256 concurrent requests.
  7. 4 NovemberNew white paperSystem One in Production: Decision Quality, Reliability and Performance on DGX Spark with vLLM and SGLangSix typed-decision models from 0.6B to 27B evaluated on one NVIDIA DGX Spark with vLLM and SGLang: Turkish accuracy, confidence calibration, paraphrase consistency, multimodal limits, engine portability, and the confidence thresholds a production decision policy needs.
  8. 4 NovemberNew CV benchmarkYOLOV5-X FP16 — Jetson Orin NanoYOLOV5-X object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  9. 4 NovemberNew CV benchmarkYOLOV5-S FP16 — Jetson Orin NanoYOLOV5-S object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  10. 4 NovemberNew CV benchmarkYOLOV5-N FP16 — Jetson Orin NanoYOLOV5-N object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  11. 4 NovemberNew CV benchmarkYOLOV5-M FP16 — Jetson Orin NanoYOLOV5-M object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  12. 4 NovemberNew CV benchmarkYOLOV5-L FP16 — Jetson Orin NanoYOLOV5-L object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 2 simultaneous cameras.
  13. 4 NovemberNew CV benchmarkYOLO11-L FP16 — Jetson Orin NanoYOLO11-L object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 2 simultaneous cameras.
  14. 3 NovemberNew CV benchmarkYOLOX-X FP16 — Jetson Orin NX 16GBYOLOX-X object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  15. 3 NovemberNew CV benchmarkYOLOX-S FP16 — Jetson Orin NX 16GBYOLOX-S object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  16. 3 NovemberNew CV benchmarkYOLOX-M FP16 — Jetson Orin NX 16GBYOLOX-M object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  17. 3 NovemberNew CV benchmarkYOLOX-L FP16 — Jetson Orin NX 16GBYOLOX-L object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 2 simultaneous cameras.
  18. 3 NovemberNew CV benchmarkYOLOV9-S FP16 — Jetson Orin NX 16GBYOLOV9-S object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  19. 3 NovemberNew CV benchmarkYOLOV9-M FP16 — Jetson Orin NX 16GBYOLOV9-M object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  20. 3 NovemberNew CV benchmarkYOLOV8-X FP16 — Jetson Orin NX 16GBYOLOV8-X object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  21. 3 NovemberNew CV benchmarkYOLOV8-S FP16 — Jetson Orin NX 16GBYOLOV8-S object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  22. 3 NovemberNew CV benchmarkYOLOV8-N FP16 — Jetson Orin NX 16GBYOLOV8-N object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  23. 3 NovemberNew CV benchmarkYOLOV8-M FP16 — Jetson Orin NX 16GBYOLOV8-M object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  24. 3 NovemberNew CV benchmarkYOLOV8-L FP16 — Jetson Orin NX 16GBYOLOV8-L object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 2 simultaneous cameras.
  25. 3 NovemberNew CV benchmarkYOLOV6-S FP16 — Jetson Orin NX 16GBYOLOV6-S object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  26. 3 NovemberNew CV benchmarkYOLOV6-N FP16 — Jetson Orin NX 16GBYOLOV6-N object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  27. 3 NovemberNew CV benchmarkYOLOV6-M FP16 — Jetson Orin NX 16GBYOLOV6-M object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  28. 3 NovemberNew CV benchmarkYOLOV6-L FP16 — Jetson Orin NX 16GBYOLOV6-L object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 2 simultaneous cameras.
  29. 3 NovemberNew CV benchmarkYOLOV5U-X FP16 — Jetson Orin NX 16GBYOLOV5U-X object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  30. 3 NovemberNew CV benchmarkYOLOV5U-S FP16 — Jetson Orin NX 16GBYOLOV5U-S object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  31. 3 NovemberNew CV benchmarkYOLOV5U-M FP16 — Jetson Orin NX 16GBYOLOV5U-M object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  32. 3 NovemberNew CV benchmarkYOLOV5U-L FP16 — Jetson Orin NX 16GBYOLOV5U-L object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  33. 3 NovemberNew CV benchmarkYOLOV5-X FP16 — Jetson Orin NX 16GBYOLOV5-X object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  34. 3 NovemberNew CV benchmarkYOLOV5-S FP16 — Jetson Orin NX 16GBYOLOV5-S object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  35. 3 NovemberNew CV benchmarkYOLOV5-N FP16 — Jetson Orin NX 16GBYOLOV5-N object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  36. 3 NovemberNew CV benchmarkYOLOV5-M FP16 — Jetson Orin NX 16GBYOLOV5-M object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  37. 3 NovemberNew CV benchmarkYOLOV5-L FP16 — Jetson Orin NX 16GBYOLOV5-L object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  38. 3 NovemberNew CV benchmarkYOLOV10-X FP16 — Jetson Orin NX 16GBYOLOV10-X object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  39. 3 NovemberNew CV benchmarkYOLOV10-S FP16 — Jetson Orin NX 16GBYOLOV10-S object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  40. 3 NovemberNew CV benchmarkYOLOV10-N FP16 — Jetson Orin NX 16GBYOLOV10-N object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  41. 3 NovemberNew CV benchmarkYOLOV10-M FP16 — Jetson Orin NX 16GBYOLOV10-M object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  42. 3 NovemberNew CV benchmarkYOLOV10-L FP16 — Jetson Orin NX 16GBYOLOV10-L object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 2 simultaneous cameras.
  43. 3 NovemberNew CV benchmarkYOLO11-X FP16 — Jetson Orin NX 16GBYOLO11-X object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  44. 3 NovemberNew CV benchmarkYOLO11-S FP16 — Jetson Orin NX 16GBYOLO11-S object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  45. 3 NovemberNew CV benchmarkYOLO11-N FP16 — Jetson Orin NX 16GBYOLO11-N object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  46. 3 NovemberNew CV benchmarkYOLO11-L FP16 — Jetson Orin NX 16GBYOLO11-L object detection (FP16, 640×640 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  47. 3 NovemberNew CV benchmarkTrafficCamNet INT8 — Jetson Orin NX 16GBTrafficCamNet vehicle detection (INT8, 960×544 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  48. 3 NovemberNew CV benchmarkPeopleNet INT8 — Jetson Orin NX 16GBPeopleNet person detection (INT8, 960×544 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  49. 3 NovemberNew CV benchmarkLPDNet INT8 — Jetson Orin NX 16GBLPDNet license plate detection (INT8, 640×480 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  50. 3 NovemberNew CV benchmarkDashCamNet INT8 — Jetson Orin NX 16GBDashCamNet vehicle detection (INT8, 960×544 input) on Jetson Orin NX 16GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  51. 3 NovemberNew CV benchmarkYOLOX-X FP16 — Jetson Orin NanoYOLOX-X object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  52. 3 NovemberNew CV benchmarkYOLOX-S FP16 — Jetson Orin NanoYOLOX-S object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  53. 3 NovemberNew CV benchmarkYOLOX-M FP16 — Jetson Orin NanoYOLOX-M object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 2 simultaneous cameras.
  54. 3 NovemberNew CV benchmarkYOLOX-L FP16 — Jetson Orin NanoYOLOX-L object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  55. 3 NovemberNew CV benchmarkYOLOV9-S FP16 — Jetson Orin NanoYOLOV9-S object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  56. 3 NovemberNew CV benchmarkYOLOV9-M FP16 — Jetson Orin NanoYOLOV9-M object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 2 simultaneous cameras.
  57. 3 NovemberNew CV benchmarkYOLOV8-S FP16 — Jetson Orin NanoYOLOV8-S object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  58. 3 NovemberNew CV benchmarkYOLOV8-N FP16 — Jetson Orin NanoYOLOV8-N object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  59. 3 NovemberNew CV benchmarkYOLOV8-M FP16 — Jetson Orin NanoYOLOV8-M object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 2 simultaneous cameras.
  60. 3 NovemberNew CV benchmarkYOLOV8-L FP16 — Jetson Orin NanoYOLOV8-L object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  61. 3 NovemberNew CV benchmarkYOLOV6-S FP16 — Jetson Orin NanoYOLOV6-S object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  62. 3 NovemberNew CV benchmarkYOLOV6-N FP16 — Jetson Orin NanoYOLOV6-N object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  63. 3 NovemberNew CV benchmarkYOLOV6-M FP16 — Jetson Orin NanoYOLOV6-M object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 2 simultaneous cameras.
  64. 3 NovemberNew CV benchmarkYOLOV6-L FP16 — Jetson Orin NanoYOLOV6-L object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 2 simultaneous cameras.
  65. 3 NovemberNew CV benchmarkYOLOV5U-X FP16 — Jetson Orin NanoYOLOV5U-X object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  66. 3 NovemberNew CV benchmarkYOLOV5U-S FP16 — Jetson Orin NanoYOLOV5U-S object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  67. 3 NovemberNew CV benchmarkYOLOV5U-M FP16 — Jetson Orin NanoYOLOV5U-M object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  68. 3 NovemberNew CV benchmarkYOLOV5U-L FP16 — Jetson Orin NanoYOLOV5U-L object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 2 simultaneous cameras.
  69. 3 NovemberNew CV benchmarkYOLOV10-X FP16 — Jetson Orin NanoYOLOV10-X object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  70. 3 NovemberNew CV benchmarkYOLOV10-S FP16 — Jetson Orin NanoYOLOV10-S object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  71. 3 NovemberNew CV benchmarkYOLOV10-N FP16 — Jetson Orin NanoYOLOV10-N object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  72. 3 NovemberNew CV benchmarkYOLOV10-M FP16 — Jetson Orin NanoYOLOV10-M object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 2 simultaneous cameras.
  73. 3 NovemberNew CV benchmarkYOLOV10-L FP16 — Jetson Orin NanoYOLOV10-L object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  74. 3 NovemberNew CV benchmarkYOLO11-X FP16 — Jetson Orin NanoYOLO11-X object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 1 simultaneous cameras.
  75. 3 NovemberNew CV benchmarkYOLO11-S FP16 — Jetson Orin NanoYOLO11-S object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  76. 3 NovemberNew CV benchmarkYOLO11-N FP16 — Jetson Orin NanoYOLO11-N object detection (FP16, 640×640 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  77. 3 NovemberNew CV benchmarkTrafficCamNet INT8 — Jetson Orin NanoTrafficCamNet vehicle detection (INT8, 960×544 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  78. 3 NovemberNew CV benchmarkPeopleNet INT8 — Jetson Orin NanoPeopleNet person detection (INT8, 960×544 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  79. 3 NovemberNew CV benchmarkLPDNet INT8 — Jetson Orin NanoLPDNet license plate detection (INT8, 640×480 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  80. 3 NovemberNew CV benchmarkDashCamNet INT8 — Jetson Orin NanoDashCamNet vehicle detection (INT8, 960×544 input) on Jetson Orin Nano with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  81. 3 NovemberNew CV benchmarkYOLOX-X FP16 — Jetson AGX Orin 64GBYOLOX-X object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 4 simultaneous cameras.
  82. 3 NovemberNew CV benchmarkYOLOX-S FP16 — Jetson AGX Orin 64GBYOLOX-S object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  83. 3 NovemberNew CV benchmarkYOLOX-M FP16 — Jetson AGX Orin 64GBYOLOX-M object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  84. 3 NovemberNew CV benchmarkYOLOX-L FP16 — Jetson AGX Orin 64GBYOLOX-L object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  85. 3 NovemberNew CV benchmarkYOLOV9-S FP16 — Jetson AGX Orin 64GBYOLOV9-S object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  86. 3 NovemberNew CV benchmarkYOLOV9-M FP16 — Jetson AGX Orin 64GBYOLOV9-M object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  87. 3 NovemberNew CV benchmarkYOLOV8-M FP16 — Jetson AGX Orin 64GBYOLOV8-M object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  88. 3 NovemberNew CV benchmarkYOLOV8-L FP16 — Jetson AGX Orin 64GBYOLOV8-L object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  89. 3 NovemberNew CV benchmarkYOLOV6-S FP16 — Jetson AGX Orin 64GBYOLOV6-S object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  90. 3 NovemberNew CV benchmarkYOLOV6-N FP16 — Jetson AGX Orin 64GBYOLOV6-N object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  91. 3 NovemberNew CV benchmarkYOLOV6-M FP16 — Jetson AGX Orin 64GBYOLOV6-M object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  92. 3 NovemberNew CV benchmarkYOLOV6-L FP16 — Jetson AGX Orin 64GBYOLOV6-L object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  93. 3 NovemberNew CV benchmarkYOLOV5U-S FP16 — Jetson AGX Orin 64GBYOLOV5U-S object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  94. 3 NovemberNew CV benchmarkYOLOV5U-M FP16 — Jetson AGX Orin 64GBYOLOV5U-M object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  95. 3 NovemberNew CV benchmarkYOLOV5U-L FP16 — Jetson AGX Orin 64GBYOLOV5U-L object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  96. 3 NovemberNew CV benchmarkYOLOV10-S FP16 — Jetson AGX Orin 64GBYOLOV10-S object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  97. 3 NovemberNew CV benchmarkYOLOV10-N FP16 — Jetson AGX Orin 64GBYOLOV10-N object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  98. 3 NovemberNew CV benchmarkYOLOV10-M FP16 — Jetson AGX Orin 64GBYOLOV10-M object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  99. 3 NovemberNew CV benchmarkYOLOV10-L FP16 — Jetson AGX Orin 64GBYOLOV10-L object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  100. 3 NovemberNew CV benchmarkYOLOX-X FP16 — GB10YOLOX-X object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  101. 3 NovemberNew CV benchmarkYOLOX-S FP16 — GB10YOLOX-S object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  102. 3 NovemberNew CV benchmarkYOLOX-M FP16 — GB10YOLOX-M object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  103. 3 NovemberNew CV benchmarkYOLOX-L FP16 — GB10YOLOX-L object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  104. 3 NovemberNew CV benchmarkYOLOV9-S FP16 — GB10YOLOV9-S object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  105. 3 NovemberNew CV benchmarkYOLOV9-M FP16 — GB10YOLOV9-M object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  106. 3 NovemberNew CV benchmarkYOLOV8-X FP16 — GB10YOLOV8-X object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  107. 3 NovemberNew CV benchmarkYOLOV8-S FP16 — GB10YOLOV8-S object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  108. 3 NovemberNew CV benchmarkYOLOV8-N FP16 — GB10YOLOV8-N object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 48 simultaneous cameras.
  109. 3 NovemberNew CV benchmarkYOLOV8-M FP16 — GB10YOLOV8-M object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  110. 3 NovemberNew CV benchmarkYOLOV8-L FP16 — GB10YOLOV8-L object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  111. 3 NovemberNew CV benchmarkYOLOV6-S FP16 — GB10YOLOV6-S object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  112. 3 NovemberNew CV benchmarkYOLOV6-N FP16 — GB10YOLOV6-N object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 48 simultaneous cameras.
  113. 3 NovemberNew CV benchmarkYOLOV6-M FP16 — GB10YOLOV6-M object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  114. 3 NovemberNew CV benchmarkYOLOV6-L FP16 — GB10YOLOV6-L object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  115. 3 NovemberNew CV benchmarkYOLOV5U-X FP16 — GB10YOLOV5U-X object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  116. 3 NovemberNew CV benchmarkYOLOV5U-S FP16 — GB10YOLOV5U-S object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 48 simultaneous cameras.
  117. 3 NovemberNew CV benchmarkYOLOV5U-M FP16 — GB10YOLOV5U-M object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  118. 3 NovemberNew CV benchmarkYOLOV5U-L FP16 — GB10YOLOV5U-L object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  119. 3 NovemberNew CV benchmarkYOLOV5-X FP16 — GB10YOLOV5-X object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  120. 3 NovemberNew CV benchmarkYOLOV5-S FP16 — GB10YOLOV5-S object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 48 simultaneous cameras.
  121. 3 NovemberNew CV benchmarkYOLOV5-N FP16 — GB10YOLOV5-N object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 48 simultaneous cameras.
  122. 3 NovemberNew CV benchmarkYOLOV5-M FP16 — GB10YOLOV5-M object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  123. 3 NovemberNew CV benchmarkYOLOV5-L FP16 — GB10YOLOV5-L object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  124. 3 NovemberNew CV benchmarkYOLOV10-X FP16 — GB10YOLOV10-X object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  125. 3 NovemberNew CV benchmarkYOLOV10-S FP16 — GB10YOLOV10-S object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  126. 3 NovemberNew CV benchmarkYOLOV10-N FP16 — GB10YOLOV10-N object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 48 simultaneous cameras.
  127. 3 NovemberNew CV benchmarkYOLOV10-M FP16 — GB10YOLOV10-M object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  128. 3 NovemberNew CV benchmarkYOLOV10-L FP16 — GB10YOLOV10-L object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  129. 3 NovemberNew CV benchmarkYOLO11-X FP16 — GB10YOLO11-X object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  130. 3 NovemberNew CV benchmarkYOLO11-S FP16 — GB10YOLO11-S object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  131. 3 NovemberNew CV benchmarkYOLO11-N FP16 — GB10YOLO11-N object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 48 simultaneous cameras.
  132. 3 NovemberNew CV benchmarkYOLO11-L FP16 — GB10YOLO11-L object detection (FP16, 640×640 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  133. 3 NovemberNew CV benchmarkLPDNet INT8 — GB10LPDNet license plate detection (INT8, 640×480 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 64 simultaneous cameras.
  134. 2 NovemberNew CV benchmarkYOLOX-X FP16 — RTX 3090YOLOX-X object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  135. 2 NovemberNew CV benchmarkYOLOX-S FP16 — RTX 3090YOLOX-S object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  136. 2 NovemberNew CV benchmarkYOLOX-M FP16 — RTX 3090YOLOX-M object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  137. 2 NovemberNew CV benchmarkYOLOX-L FP16 — RTX 3090YOLOX-L object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  138. 2 NovemberNew CV benchmarkYOLOV9-S FP16 — RTX 3090YOLOV9-S object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  139. 2 NovemberNew CV benchmarkYOLOV9-M FP16 — RTX 3090YOLOV9-M object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  140. 2 NovemberNew CV benchmarkYOLOV8-X FP16 — RTX 3090YOLOV8-X object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  141. 2 NovemberNew CV benchmarkYOLOV8-S FP16 — RTX 3090YOLOV8-S object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  142. 2 NovemberNew CV benchmarkYOLOV8-N FP16 — RTX 3090YOLOV8-N object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  143. 2 NovemberNew CV benchmarkYOLOV8-M FP16 — RTX 3090YOLOV8-M object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  144. 2 NovemberNew CV benchmarkYOLOV8-L FP16 — RTX 3090YOLOV8-L object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  145. 2 NovemberNew CV benchmarkYOLOV6-S FP16 — RTX 3090YOLOV6-S object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  146. 2 NovemberNew CV benchmarkYOLOV6-N FP16 — RTX 3090YOLOV6-N object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  147. 2 NovemberNew CV benchmarkYOLOV6-M FP16 — RTX 3090YOLOV6-M object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  148. 2 NovemberNew CV benchmarkYOLOV6-L FP16 — RTX 3090YOLOV6-L object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  149. 2 NovemberNew CV benchmarkYOLOV5U-X FP16 — RTX 3090YOLOV5U-X object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  150. 2 NovemberNew CV benchmarkYOLOV5U-S FP16 — RTX 3090YOLOV5U-S object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  151. 2 NovemberNew CV benchmarkYOLOV5U-M FP16 — RTX 3090YOLOV5U-M object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  152. 2 NovemberNew CV benchmarkYOLOV5U-L FP16 — RTX 3090YOLOV5U-L object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  153. 2 NovemberNew CV benchmarkYOLOV5-X FP16 — RTX 3090YOLOV5-X object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  154. 2 NovemberNew CV benchmarkYOLOV5-S FP16 — RTX 3090YOLOV5-S object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  155. 2 NovemberNew CV benchmarkYOLOV5-N FP16 — RTX 3090YOLOV5-N object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  156. 2 NovemberNew CV benchmarkYOLOV5-M FP16 — RTX 3090YOLOV5-M object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  157. 2 NovemberNew CV benchmarkYOLOV5-L FP16 — RTX 3090YOLOV5-L object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  158. 2 NovemberNew CV benchmarkYOLOV10-X FP16 — RTX 3090YOLOV10-X object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  159. 2 NovemberNew CV benchmarkYOLOV10-S FP16 — RTX 3090YOLOV10-S object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  160. 2 NovemberNew CV benchmarkYOLOV10-N FP16 — RTX 3090YOLOV10-N object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  161. 2 NovemberNew CV benchmarkYOLOV10-M FP16 — RTX 3090YOLOV10-M object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  162. 2 NovemberNew CV benchmarkYOLOV10-L FP16 — RTX 3090YOLOV10-L object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  163. 2 NovemberNew CV benchmarkYOLO11-X FP16 — RTX 3090YOLO11-X object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  164. 2 NovemberNew CV benchmarkYOLO11-S FP16 — RTX 3090YOLO11-S object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  165. 2 NovemberNew CV benchmarkYOLO11-N FP16 — RTX 3090YOLO11-N object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  166. 2 NovemberNew CV benchmarkYOLO11-L FP16 — RTX 3090YOLO11-L object detection (FP16, 640×640 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  167. 2 NovemberNew CV benchmarkTrafficCamNet INT8 — RTX 3090TrafficCamNet vehicle detection (INT8, 960×544 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  168. 2 NovemberNew CV benchmarkPeopleNet INT8 — RTX 3090PeopleNet person detection (INT8, 960×544 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 42 simultaneous cameras.
  169. 2 NovemberNew CV benchmarkLPDNet INT8 — RTX 3090LPDNet license plate detection (INT8, 640×480 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  170. 2 NovemberNew CV benchmarkDashCamNet INT8 — RTX 3090DashCamNet vehicle detection (INT8, 960×544 input) on RTX 3090 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 40 simultaneous cameras.
  171. 2 NovemberNew CV benchmarkYOLOV8-S FP16 — Jetson AGX Orin 64GBYOLOV8-S object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  172. 2 NovemberNew CV benchmarkYOLOV8-N FP16 — Jetson AGX Orin 64GBYOLOV8-N object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  173. 2 NovemberNew CV benchmarkYOLO11-S FP16 — Jetson AGX Orin 64GBYOLO11-S object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  174. 2 NovemberNew CV benchmarkYOLO11-N FP16 — Jetson AGX Orin 64GBYOLO11-N object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  175. 2 NovemberNew CV benchmarkYOLO11-L FP16 — Jetson AGX Orin 64GBYOLO11-L object detection (FP16, 640×640 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 8 simultaneous cameras.
  176. 2 NovemberNew CV benchmarkTrafficCamNet INT8 — Jetson AGX Orin 64GBTrafficCamNet vehicle detection (INT8, 960×544 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  177. 2 NovemberNew CV benchmarkPeopleNet INT8 — Jetson AGX Orin 64GBPeopleNet person detection (INT8, 960×544 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  178. 2 NovemberNew CV benchmarkLPDNet INT8 — Jetson AGX Orin 64GBLPDNet license plate detection (INT8, 640×480 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 48 simultaneous cameras.
  179. 2 NovemberNew CV benchmarkDashCamNet INT8 — Jetson AGX Orin 64GBDashCamNet vehicle detection (INT8, 960×544 input) on Jetson AGX Orin 64GB with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  180. 2 NovemberNew CV benchmarkTrafficCamNet INT8 — GB10TrafficCamNet vehicle detection (INT8, 960×544 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 48 simultaneous cameras.
  181. 2 NovemberNew CV benchmarkPeopleNet INT8 — GB10PeopleNet person detection (INT8, 960×544 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 48 simultaneous cameras.
  182. 2 NovemberNew CV benchmarkDashCamNet INT8 — GB10DashCamNet vehicle detection (INT8, 960×544 input) on GB10 with a replayed 1080p20 clip: frames per second per camera, measured from 1 to 48 simultaneous cameras.
  183. 2 NovemberNew LLM benchmarkMiniMax-M3.1-Flash — DGX B300MiniMax-M3.1-Flash on DGX B300: MXFP8 dense layers and NVFP4 expert weights served with vLLM at TP=2, with DSpark speculative decoding (k=8) and a 1M context, measured up to 256 concurrent requests.
  184. 2 NovemberNew LLM benchmarkMiniMax-M3.1-Flash — 4× DGX SparkMiniMax-M3.1-Flash on 4× DGX Spark: MXFP8 dense layers and NVFP4 expert weights served with vLLM at TP=4, with DSpark speculative decoding (k=8) and a 1M context, measured up to 8 concurrent requests.
  185. 2 NovemberNew LLM benchmarkNVIDIA-Nemotron-3.5-Bolt-74B-A9B — DGX B300NVIDIA-Nemotron-3.5-Bolt-74B-A9B on DGX B300: BF16 weights served with vLLM at TP=1, with a 256K context, measured up to 256 concurrent requests.

October 2026

  1. 28 OctoberNew LLM benchmarkDeepSeek-V4.1-Flash-1017 — DGX B300DeepSeek-V4.1-Flash-1017 on DGX B300: FP8 weights served with vLLM at TP=2, with DSpark speculative decoding (k=3) and a 1M context, measured up to 256 concurrent requests.
  2. 26 OctoberNew LLM benchmarkMuse-Spark-Open-816B-A42B — 8× DGX SparkMuse-Spark-Open-816B-A42B on 8× DGX Spark: NVFP4 weights served with vLLM at TP=8, with DSpark speculative decoding and a 128K context, measured up to 8 concurrent requests.
  3. 23 OctoberNew VLM benchmarkd1-3B — DGX Sparkd1-3B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 256 simultaneous cameras and up to 32 images per request.
  4. 23 OctoberNew VLM benchmarkFara1.5-27B — DGX SparkFara1.5-27B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 40 simultaneous cameras and up to 128 images per request.
  5. 23 OctoberNew VLM benchmarkGemma-4-26B-A4B-it — DGX SparkGemma-4-26B-A4B-it on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 120 simultaneous cameras and up to 128 images per request.
  6. 23 OctoberNew VLM benchmarkGLM-4.6V-Flash — DGX SparkGLM-4.6V-Flash on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 48 simultaneous cameras and up to 64 images per request.
  7. 23 OctoberNew VLM benchmarkGranite-Vision-4.1-4B — DGX SparkGranite-Vision-4.1-4B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 96 simultaneous cameras and up to 32 images per request.
  8. 23 OctoberNew VLM benchmarkH2O-Lightning-4B — DGX SparkH2O-Lightning-4B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 144 simultaneous cameras and up to 128 images per request.
  9. 23 OctoberNew VLM benchmarkLensVLM-9B — DGX SparkLensVLM-9B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 80 simultaneous cameras and up to 128 images per request.
  10. 23 OctoberNew VLM benchmarkLFM2.5-VL-3B — DGX SparkLFM2.5-VL-3B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 256 simultaneous cameras and up to 32 images per request.
  11. 23 OctoberNew VLM benchmarkLlama-4-Scout-17B-16E-Instruct — DGX SparkLlama-4-Scout-17B-16E-Instruct on DGX Spark: NVFP4 weights served with vLLM, response time measured from 1 to 28 simultaneous cameras and up to 64 images per request.
  12. 23 OctoberNew VLM benchmarkMiMo-V2.6-Distill-Qwen-9B — DGX SparkMiMo-V2.6-Distill-Qwen-9B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 96 simultaneous cameras and up to 128 images per request.
  13. 23 OctoberNew VLM benchmarkMiniCPM-V-4.6 — DGX SparkMiniCPM-V-4.6 on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 384 simultaneous cameras and up to 128 images per request.
  14. 23 OctoberNew VLM benchmarkMiniCPM-V-4.7-1B — DGX SparkMiniCPM-V-4.7-1B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 352 simultaneous cameras and up to 128 images per request.
  15. 23 OctoberNew VLM benchmarkMistral-Small-4-119B-2603 — DGX SparkMistral-Small-4-119B-2603 on DGX Spark: NVFP4 weights served with vLLM, response time measured from 1 to 88 simultaneous cameras and up to 64 images per request.
  16. 23 OctoberNew VLM benchmarkNVIDIA-Nemotron-3.5-Bolt-74B-A9B — DGX SparkNVIDIA-Nemotron-3.5-Bolt-74B-A9B on DGX Spark: NVFP4 weights served with vLLM, response time measured from 1 to 64 simultaneous cameras and up to 32 images per request.
  17. 23 OctoberNew VLM benchmarkNVIDIA-Nemotron-3.5-Super-VL-120B-A12B — DGX SparkNVIDIA-Nemotron-3.5-Super-VL-120B-A12B on DGX Spark: NVFP4 weights served with vLLM, response time measured from 1 to 56 simultaneous cameras and up to 64 images per request.
  18. 23 OctoberNew VLM benchmarkQwen3.5-0.8B — DGX SparkQwen3.5-0.8B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 416 simultaneous cameras and up to 128 images per request.
  19. 23 OctoberNew VLM benchmarkQwen3.5-122B-A10B — DGX SparkQwen3.5-122B-A10B on DGX Spark: NVFP4 weights served with vLLM, response time measured from 1 to 20 simultaneous cameras and up to 128 images per request.
  20. 23 OctoberNew VLM benchmarkQwen3.5-27B — DGX SparkQwen3.5-27B on DGX Spark: FP8 weights served with vLLM, response time measured from 1 to 56 simultaneous cameras and up to 128 images per request.
  21. 23 OctoberNew VLM benchmarkQwen3.5-2B — DGX SparkQwen3.5-2B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 240 simultaneous cameras and up to 128 images per request.
  22. 23 OctoberNew VLM benchmarkQwen3.5-35B-A3B — DGX SparkQwen3.5-35B-A3B on DGX Spark: FP8 weights served with vLLM, response time measured from 1 to 112 simultaneous cameras and up to 128 images per request.
  23. 23 OctoberNew VLM benchmarkQwen3.5-9B — DGX SparkQwen3.5-9B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 96 simultaneous cameras and up to 128 images per request.
  24. 23 OctoberNew VLM benchmarkQwen3.6-27B — DGX SparkQwen3.6-27B on DGX Spark: FP8 weights served with vLLM, response time measured from 1 to 56 simultaneous cameras and up to 128 images per request.
  25. 23 OctoberNew VLM benchmarkd1-3B — Jetson Orin NX 16GBd1-3B on Jetson Orin NX 16GB: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  26. 23 OctoberNew VLM benchmarkGLM-4.6V-Flash — Jetson Orin NX 16GBGLM-4.6V-Flash on Jetson Orin NX 16GB: Q4_K_M GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  27. 23 OctoberNew VLM benchmarkGranite-Vision-4.1-4B — Jetson Orin NX 16GBGranite-Vision-4.1-4B on Jetson Orin NX 16GB: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  28. 23 OctoberNew VLM benchmarkLensVLM-9B — Jetson Orin NX 16GBLensVLM-9B on Jetson Orin NX 16GB: Q4_K_M GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  29. 23 OctoberNew VLM benchmarkLFM2.5-VL-3B — Jetson Orin NX 16GBLFM2.5-VL-3B on Jetson Orin NX 16GB: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  30. 23 OctoberNew VLM benchmarkMiMo-V2.6-Distill-Qwen-9B — Jetson Orin NX 16GBMiMo-V2.6-Distill-Qwen-9B on Jetson Orin NX 16GB: Q4_K_M GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  31. 23 OctoberNew VLM benchmarkMiniCPM-V-4.6 — Jetson Orin NX 16GBMiniCPM-V-4.6 on Jetson Orin NX 16GB: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  32. 23 OctoberNew VLM benchmarkMiniCPM-V-4.7-1B — Jetson Orin NX 16GBMiniCPM-V-4.7-1B on Jetson Orin NX 16GB: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  33. 23 OctoberNew VLM benchmarkQwen-Drive-1.0-4B — Jetson Orin NX 16GBQwen-Drive-1.0-4B on Jetson Orin NX 16GB: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  34. 23 OctoberNew VLM benchmarkQwen3.5-0.8B — Jetson Orin NX 16GBQwen3.5-0.8B on Jetson Orin NX 16GB: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  35. 23 OctoberNew VLM benchmarkQwen3.5-2B — Jetson Orin NX 16GBQwen3.5-2B on Jetson Orin NX 16GB: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  36. 23 OctoberNew VLM benchmarkQwen3.5-9B — Jetson Orin NX 16GBQwen3.5-9B on Jetson Orin NX 16GB: Q4_K_M GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  37. 23 OctoberNew VLM benchmarkd1-3B — Jetson AGX Orin 32GB Developer Kitd1-3B on Jetson AGX Orin 32GB Developer Kit: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  38. 23 OctoberNew VLM benchmarkGLM-4.6V-Flash — Jetson AGX Orin 32GB Developer KitGLM-4.6V-Flash on Jetson AGX Orin 32GB Developer Kit: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  39. 23 OctoberNew VLM benchmarkGranite-Vision-4.1-4B — Jetson AGX Orin 32GB Developer KitGranite-Vision-4.1-4B on Jetson AGX Orin 32GB Developer Kit: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  40. 23 OctoberNew VLM benchmarkLensVLM-9B — Jetson AGX Orin 32GB Developer KitLensVLM-9B on Jetson AGX Orin 32GB Developer Kit: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  41. 23 OctoberNew VLM benchmarkLFM2.5-VL-3B — Jetson AGX Orin 32GB Developer KitLFM2.5-VL-3B on Jetson AGX Orin 32GB Developer Kit: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  42. 23 OctoberNew VLM benchmarkMiMo-V2.6-Distill-Qwen-9B — Jetson AGX Orin 32GB Developer KitMiMo-V2.6-Distill-Qwen-9B on Jetson AGX Orin 32GB Developer Kit: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  43. 23 OctoberNew VLM benchmarkMiniCPM-V-4.6 — Jetson AGX Orin 32GB Developer KitMiniCPM-V-4.6 on Jetson AGX Orin 32GB Developer Kit: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  44. 23 OctoberNew VLM benchmarkMiniCPM-V-4.7-1B — Jetson AGX Orin 32GB Developer KitMiniCPM-V-4.7-1B on Jetson AGX Orin 32GB Developer Kit: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  45. 23 OctoberNew VLM benchmarkMiniCPM-V-4.7-35B-A3B — Jetson AGX Orin 32GB Developer KitMiniCPM-V-4.7-35B-A3B on Jetson AGX Orin 32GB Developer Kit: Q4_K_M GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  46. 23 OctoberNew VLM benchmarkQwen-Drive-1.0-4B — Jetson AGX Orin 32GB Developer KitQwen-Drive-1.0-4B on Jetson AGX Orin 32GB Developer Kit: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  47. 23 OctoberNew VLM benchmarkQwen3.5-0.8B — Jetson AGX Orin 32GB Developer KitQwen3.5-0.8B on Jetson AGX Orin 32GB Developer Kit: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  48. 23 OctoberNew VLM benchmarkQwen3.5-27B — Jetson AGX Orin 32GB Developer KitQwen3.5-27B on Jetson AGX Orin 32GB Developer Kit: Q4_K_M GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  49. 23 OctoberNew VLM benchmarkQwen3.5-2B — Jetson AGX Orin 32GB Developer KitQwen3.5-2B on Jetson AGX Orin 32GB Developer Kit: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  50. 23 OctoberNew VLM benchmarkQwen3.5-35B-A3B — Jetson AGX Orin 32GB Developer KitQwen3.5-35B-A3B on Jetson AGX Orin 32GB Developer Kit: Q4_K_M GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  51. 23 OctoberNew VLM benchmarkQwen3.5-9B — Jetson AGX Orin 32GB Developer KitQwen3.5-9B on Jetson AGX Orin 32GB Developer Kit: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  52. 23 OctoberNew VLM benchmarkQwen3.6-27B — Jetson AGX Orin 32GB Developer KitQwen3.6-27B on Jetson AGX Orin 32GB Developer Kit: Q4_K_M GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 32 images per request.
  53. 21 OctoberNew LLM benchmarkMuse-Spark-Open-816B-A42B — DGX B300Muse-Spark-Open-816B-A42B on a DGX B300: NVFP4 weights served with vLLM at TP=4, with DSpark speculative decoding, measured up to 64 concurrent requests.
  54. 21 OctoberNew LLM benchmarkNVIDIA-Nemotron-3.5-Super-VL-120B-A12B — DGX B300NVIDIA-Nemotron-3.5-Super-VL-120B-A12B on a DGX B300: BF16 weights served with vLLM at TP=2, with native MTP speculative decoding (k=3), measured up to 256 concurrent requests.
  55. 21 OctoberNew LLM benchmarkStep-5-Preview — DGX B300Step-5-Preview on a DGX B300: FP8 weights served with vLLM at TP=4, with MTP speculative decoding (k=3), measured up to 256 concurrent requests.
  56. 21 OctoberNew LLM benchmarkStep-5-Preview — DGX B300Step-5-Preview on a DGX B300 with expert parallelism: FP8 weights served with vLLM at EP=4 over TP=4, with MTP speculative decoding (k=3), measured up to 256 concurrent requests.
  57. 21 OctoberNew LLM benchmarkStep-5-Preview — DGX B300Step-5-Preview on a DGX B300 with data parallelism: FP8 weights served with vLLM at DP=4, with MTP speculative decoding (k=3), measured up to 256 concurrent requests.
  58. 21 OctoberNew white paperKV Cache Offloading with LMCache in LLM InferenceWhat LMCache adds to vLLM when the same contexts come back: a three-tier KV cache (GPU, CPU, NVMe) on DGX Spark, measured TTFT — 1.95× on repeated access, 3.3× after a service restart — and the settings and limits that come with it.
  59. 15 OctoberNew white paperNetwork Link Speed for Tensor-Parallel Inference on 8× DGX Spark (200G vs 100G)How much network bandwidth an 8-Spark tensor-parallel cluster really uses, and what running each DGX Spark at 100G instead of 200G costs: the mechanism, an estimate from the model's architecture, and measurements with GLM-5.3 at TP8.
  60. 15 OctoberNew VLM benchmarkCosmos-Reason1-7B — Jetson AGX ThorCosmos-Reason1-7B on Jetson AGX Thor: BF16 weights served with vLLM, response time measured from 1 to 16 simultaneous cameras and up to 64 images per request.
  61. 15 OctoberNew VLM benchmarkCosmos-Reason2-2B — Jetson AGX ThorCosmos-Reason2-2B on Jetson AGX Thor: BF16 weights served with vLLM, response time measured from 1 to 256 simultaneous cameras and up to 256 images per request.
  62. 15 OctoberNew VLM benchmarkCosmos-Reason2-8B — Jetson AGX ThorCosmos-Reason2-8B on Jetson AGX Thor: BF16 weights served with vLLM, response time measured from 1 to 112 simultaneous cameras and up to 256 images per request.
  63. 15 OctoberNew VLM benchmarkCosmos3-Edge — Jetson AGX ThorCosmos3-Edge on Jetson AGX Thor: BF16 weights served with vLLM, response time measured from 1 to 208 simultaneous cameras and up to 128 images per request.
  64. 15 OctoberNew VLM benchmarkCosmos3-Nano — Jetson AGX ThorCosmos3-Nano on Jetson AGX Thor: BF16 weights served with vLLM, response time measured from 1 to 112 simultaneous cameras and up to 256 images per request.
  65. 15 OctoberNew VLM benchmarkEagle2.5-8B — Jetson AGX ThorEagle2.5-8B on Jetson AGX Thor: BF16 weights served with vLLM, response time measured from 1 to 52 simultaneous cameras and up to 64 images per request.
  66. 15 OctoberNew VLM benchmarkGemma-4-31B-it — Jetson AGX ThorGemma-4-31B-it on Jetson AGX Thor: BF16 weights served with vLLM, response time measured from 1 to 60 simultaneous cameras and up to 64 images per request.
  67. 15 OctoberNew VLM benchmarkGemma-4-31B-it — Jetson AGX ThorGemma-4-31B-it on Jetson AGX Thor: NVFP4 weights served with vLLM, response time measured from 1 to 80 simultaneous cameras and up to 64 images per request.
  68. 15 OctoberNew VLM benchmarkLlama-3.1-Nemotron-Nano-VL-8B — Jetson AGX ThorLlama-3.1-Nemotron-Nano-VL-8B on Jetson AGX Thor: BF16 weights served with vLLM, response time measured from 1 to 40 simultaneous cameras and up to 32 images per request.
  69. 15 OctoberNew VLM benchmarkMedGemma-1.5-4B-it — Jetson AGX ThorMedGemma-1.5-4B-it on Jetson AGX Thor: BF16 weights served with vLLM, response time measured from 1 to 192 simultaneous cameras and up to 256 images per request.
  70. 15 OctoberNew VLM benchmarkNemotron-Nano-12B-v2-VL — Jetson AGX ThorNemotron-Nano-12B-v2-VL on Jetson AGX Thor: BF16 weights served with vLLM, response time measured from 1 to 88 simultaneous cameras and up to 128 images per request.
  71. 15 OctoberNew VLM benchmarkQwen3-VL-30B-A3B-Instruct — Jetson AGX ThorQwen3-VL-30B-A3B-Instruct on Jetson AGX Thor: FP8 weights served with vLLM, response time measured from 1 to 96 simultaneous cameras and up to 256 images per request.
  72. 15 OctoberNew VLM benchmarkQwen3-VL-4B-Instruct — Jetson AGX ThorQwen3-VL-4B-Instruct on Jetson AGX Thor: BF16 weights served with vLLM, response time measured from 1 to 160 simultaneous cameras and up to 256 images per request.
  73. 15 OctoberNew VLM benchmarkQwen3-VL-8B-Instruct — Jetson AGX ThorQwen3-VL-8B-Instruct on Jetson AGX Thor: BF16 weights served with vLLM, response time measured from 1 to 104 simultaneous cameras and up to 256 images per request.
  74. 15 OctoberNew VLM benchmarkQwen3.6-35B-A3B — Jetson AGX ThorQwen3.6-35B-A3B on Jetson AGX Thor: FP8 weights served with vLLM, response time measured from 1 to 104 simultaneous cameras and up to 256 images per request.
  75. 15 OctoberNew VLM benchmarkQwen3.6-35B-A3B — Jetson AGX ThorQwen3.6-35B-A3B on Jetson AGX Thor: NVFP4 weights served with vLLM, response time measured from 1 to 104 simultaneous cameras and up to 256 images per request.
  76. 15 OctoberNew VLM benchmarkQwen3.8-27B — Jetson AGX ThorQwen3.8-27B on Jetson AGX Thor: FP8 weights served with vLLM, response time measured from 1 to 44 simultaneous cameras and up to 256 images per request.
  77. 15 OctoberNew VLM benchmarkCosmos-Reason1-7B — DGX SparkCosmos-Reason1-7B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 16 simultaneous cameras and up to 64 images per request.
  78. 15 OctoberNew VLM benchmarkCosmos-Reason2-2B — DGX SparkCosmos-Reason2-2B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 224 simultaneous cameras and up to 256 images per request.
  79. 15 OctoberNew VLM benchmarkCosmos-Reason2-8B — DGX SparkCosmos-Reason2-8B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 96 simultaneous cameras and up to 256 images per request.
  80. 15 OctoberNew VLM benchmarkCosmos3-Edge — DGX SparkCosmos3-Edge on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 16 simultaneous cameras and up to 128 images per request.
  81. 15 OctoberNew VLM benchmarkCosmos3-Nano — DGX SparkCosmos3-Nano on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 96 simultaneous cameras and up to 256 images per request.
  82. 15 OctoberNew VLM benchmarkEagle2.5-8B — DGX SparkEagle2.5-8B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 44 simultaneous cameras and up to 64 images per request.
  83. 15 OctoberNew VLM benchmarkGemma-4-31B-it — DGX SparkGemma-4-31B-it on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 64 simultaneous cameras and up to 128 images per request.
  84. 15 OctoberNew VLM benchmarkGemma-4-31B-it — DGX SparkGemma-4-31B-it on DGX Spark: NVFP4 weights served with vLLM, response time measured from 1 to 96 simultaneous cameras and up to 128 images per request.
  85. 15 OctoberNew VLM benchmarkLlama-3.1-Nemotron-Nano-VL-8B — DGX SparkLlama-3.1-Nemotron-Nano-VL-8B on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 32 simultaneous cameras and up to 32 images per request.
  86. 15 OctoberNew VLM benchmarkMedGemma-1.5-4B-it — DGX SparkMedGemma-1.5-4B-it on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 240 simultaneous cameras and up to 256 images per request.
  87. 15 OctoberNew VLM benchmarkNemotron-Nano-12B-v2-VL — DGX SparkNemotron-Nano-12B-v2-VL on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 72 simultaneous cameras and up to 128 images per request.
  88. 15 OctoberNew VLM benchmarkQwen3-VL-30B-A3B-Instruct — DGX SparkQwen3-VL-30B-A3B-Instruct on DGX Spark: FP8 weights served with vLLM, response time measured from 1 to 104 simultaneous cameras and up to 256 images per request.
  89. 15 OctoberNew VLM benchmarkQwen3-VL-4B-Instruct — DGX SparkQwen3-VL-4B-Instruct on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 128 simultaneous cameras and up to 256 images per request.
  90. 15 OctoberNew VLM benchmarkQwen3-VL-8B-Instruct — DGX SparkQwen3-VL-8B-Instruct on DGX Spark: BF16 weights served with vLLM, response time measured from 1 to 88 simultaneous cameras and up to 256 images per request.
  91. 15 OctoberNew VLM benchmarkQwen3.6-35B-A3B — DGX SparkQwen3.6-35B-A3B on DGX Spark: FP8 weights served with vLLM, response time measured from 1 to 104 simultaneous cameras and up to 256 images per request.
  92. 15 OctoberNew VLM benchmarkQwen3.6-35B-A3B — DGX SparkQwen3.6-35B-A3B on DGX Spark: NVFP4 weights served with vLLM, response time measured from 1 to 120 simultaneous cameras and up to 256 images per request.
  93. 15 OctoberNew VLM benchmarkQwen3.8-27B — DGX SparkQwen3.8-27B on DGX Spark: FP8 weights served with vLLM, response time measured from 1 to 48 simultaneous cameras and up to 256 images per request.
  94. 15 OctoberNew VLM benchmarkCosmos-Reason1-7B — Jetson AGX Orin 32GBCosmos-Reason1-7B on Jetson AGX Orin 32GB: Q4_K_M GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 5 images per request.
  95. 15 OctoberNew VLM benchmarkCosmos-Reason2-2B — Jetson AGX Orin 32GBCosmos-Reason2-2B on Jetson AGX Orin 32GB: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 5 images per request.
  96. 15 OctoberNew VLM benchmarkCosmos-Reason2-8B — Jetson AGX Orin 32GBCosmos-Reason2-8B on Jetson AGX Orin 32GB: Q4_K_M GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 5 images per request.
  97. 15 OctoberNew VLM benchmarkCosmos3-Edge — Jetson AGX Orin 32GBCosmos3-Edge on Jetson AGX Orin 32GB: BF16 weights served with vLLM, response time measured from 1 to 8 simultaneous cameras and up to 5 images per request.
  98. 15 OctoberNew VLM benchmarkGemma-4-26B-A4B-it — Jetson AGX Orin 32GBGemma-4-26B-A4B-it on Jetson AGX Orin 32GB: Q4_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 5 images per request.
  99. 15 OctoberNew VLM benchmarkGemma-4-E4B-it — Jetson AGX Orin 32GBGemma-4-E4B-it on Jetson AGX Orin 32GB: Q4_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 5 images per request.
  100. 15 OctoberNew VLM benchmarkQwen3-VL-4B-Instruct — Jetson AGX Orin 32GBQwen3-VL-4B-Instruct on Jetson AGX Orin 32GB: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 5 images per request.
  101. 15 OctoberNew VLM benchmarkQwen3-VL-8B-Instruct — Jetson AGX Orin 32GBQwen3-VL-8B-Instruct on Jetson AGX Orin 32GB: Q4_K_M GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 5 images per request.
  102. 15 OctoberNew VLM benchmarkQwen3-VL-8B-Instruct — Jetson AGX Orin 32GBQwen3-VL-8B-Instruct on Jetson AGX Orin 32GB: Q8_0 GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 5 images per request.
  103. 15 OctoberNew VLM benchmarkQwen3.5-4B — Jetson AGX Orin 32GBQwen3.5-4B on Jetson AGX Orin 32GB: Q4_K_M GGUF weights served with llama.cpp, response time measured from 1 to 8 simultaneous cameras and up to 5 images per request.
  104. 12 OctoberNew white paperLLM Quantization Selection GuideDecision guide for LLM quantization: how workload, quantization format, GPU architecture and checkpoint reliability decide the choice — with the selection flow, frequent mistakes, case examples and a quick command reference.
  105. 8 OctoberNew LLM benchmarkGLM-5.3-Flash — 4× DGX SparkGLM-5.3-Flash (321B) on 4× DGX Spark: FP8 weights served with SGLang across four nodes (TP4), with DFlash speculative decoding (k=7) and a 256K context.
  106. 8 OctoberNew LLM benchmarkMiMo-V2.6-Pro-RL — 8× DGX SparkMiMo-V2.6-Pro-RL (1.02T) on 8× DGX Spark: FP4 weights served with SGLang across eight nodes (TP8), with DFlash speculative decoding (k=7) and a 256K context.
  107. 7 OctoberNew white paperDevice AdvisorDescribe your cameras and pick a vision-language model: the page names the smallest measured device that answers every camera within your target, then plays the resulting stream back as a live timeline — how often each camera is answered, and how many of its frames the model ever sees.
  108. 6 OctoberNew LLM benchmarkQwen3.8-27B — RTX PRO 6000Qwen3.8-27B on a single RTX PRO 6000: NVFP4 weights served with vLLM, with MTP speculative decoding (k=3), measured up to 128 concurrent requests.
  109. 6 OctoberNew LLM benchmarkQwen3.8-27B — RTX PRO 6000Qwen3.8-27B on a single RTX PRO 6000: BF16 weights served with vLLM, with MTP speculative decoding (k=3), measured up to 128 concurrent requests.
  110. 6 OctoberNew white paperThe First Open-Weight Alternatives to JevSix open-weight typed-decision models compared on one NVIDIA DGX Spark (GB10) through a shared state + questions interface and one 25-question set: accuracy, calibration, option-order sensitivity and latency — and which model fits which job.
  111. 6 OctoberNew white paperVLM Inference Benchmark ExplorerExplore vision-language model inference benchmarks on RTX PRO 6000 Blackwell, Jetson AGX Orin and Jetson Orin NX. Pick the image size, set how long a camera may wait for its answer, and see how many cameras each configuration keeps up with.
  112. 1 OctoberNew LLM benchmarkGemma-4-31B-it — ThorGemma-4-31B-it on Jetson AGX Thor: NVFP4 weights served with vLLM, with MTP speculative decoding (k=3) and a 256K context.
  113. 1 OctoberNew LLM benchmarkQwen3-VL-30B-A3B-Instruct — ThorQwen3-VL-30B-A3B-Instruct, a vision-language model, on Jetson AGX Thor: FP8 weights served with vLLM, with speculative decoding and a 256K context.
  114. 1 OctoberNew LLM benchmarkQwen3.8-27B — RTX PRO 6000Qwen3.8-27B on a single RTX PRO 6000: FP8 weights served with vLLM, with MTP speculative decoding (k=3), measured up to 128 concurrent requests.
  115. 1 OctoberNew LLM benchmarkMiMo-V2.6-Pro-MOPD — DGX B300MiMo-V2.6-Pro-MOPD (1.02T) on DGX B300: FP4 weights served with vLLM on four GPUs (TP4), with DFlash speculative decoding (k=7) and a 1M-token context, measured up to 256 concurrent requests.
  116. 1 OctoberNew LLM benchmarkMiMo-V2.6-Flash-RL — 2× DGX SparkMiMo-V2.6-Flash-RL (310B) on 2× DGX Spark: FP4 weights served with vLLM across both nodes (TP2), with DFlash speculative decoding (k=7) and a 256K context.
  117. 1 OctoberNew LLM benchmarkMiMo-V2.6-Flash-RL — 4× DGX SparkMiMo-V2.6-Flash-RL (310B) on 4× DGX Spark: FP4 weights served with vLLM across four nodes (TP4), with DFlash speculative decoding (k=7) and a 500K context.

September 2026

  1. 29 SeptemberNew white paperEngram Offloading in LLM InferenceEngram tables are large but read only a few rows per token, so they can move from GPU memory to host RAM or NVMe. What that means on DGX Spark, RTX PRO 6000 and DGX B300, and measured results with DeepSeek-V4.1-Flash and Qwen3.8-Flash-Next.
  2. 28 SeptemberNew LLM benchmarkQwen3.8-Flash-Next — 1× DGX SparkQwen3.8-Flash-Next (176B) on a single DGX Spark: NVFP4 weights served with SGLang, with MTP speculative decoding (k=3) and its 47.7 GiB n-gram embedding table read from NVMe.
  3. 28 SeptemberNew LLM benchmarkQwen3.8-Flash-Next — RTX PRO 6000Qwen3.8-Flash-Next (176B) on a single RTX PRO 6000: NVFP4 weights served with SGLang, with MTP speculative decoding (k=3) and its 47.7 GiB n-gram embedding table held in host RAM.
  4. 28 SeptemberNew LLM benchmarkMiMo-V2.6-Pro-RL — DGX B300MiMo-V2.6-Pro-RL (1.02T) on DGX B300: FP4 weights served with vLLM on four GPUs (TP4), with DFlash speculative decoding (k=7) and a 1M-token context.
  5. 18 SeptemberNew CV benchmarkPeopleNet INT8 — Jetson AGX ThorPeopleNet person detection (INT8, 960×544 input) on Jetson AGX Thor: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  6. 18 SeptemberNew CV benchmarkYOLO11-N FP16 — Jetson AGX ThorYOLO11-N object detection (640×640 input) on Jetson AGX Thor: frames per second per camera, measured from 1 to 30 simultaneous cameras.
  7. 18 SeptemberNew CV benchmarkYOLO11-S FP16 — Jetson AGX ThorYOLO11-S object detection (640×640 input) on Jetson AGX Thor: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  8. 18 SeptemberNew CV benchmarkYOLO11-M FP16 — Jetson AGX ThorYOLO11-M object detection (640×640 input) on Jetson AGX Thor: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  9. 18 SeptemberNew CV benchmarkYOLO11-L FP16 — Jetson AGX ThorYOLO11-L object detection (640×640 input) on Jetson AGX Thor: frames per second per camera, measured from 1 to 32 simultaneous cameras.
  10. 17 SeptemberNew white paperDeepSeek-V4.1-Flash 8× DGX Spark TP8 DeploymentDeployment of DeepSeek-V4.1-Flash (763B MoE, FP8, DSpark k=5) on 8× NVIDIA DGX Spark (GB10) with TP8: two configurations (300K Engram-in-memory, 1M Engram-on-disk), NCCL optimization, benchmark results and TP4 comparison.
  11. 17 SeptemberNew LLM benchmarkDeepSeek-V4.1-Flash — 8× DGX SparkDeepSeek-V4.1-Flash (763B) on 8× DGX Spark: FP8 weights served with vLLM across eight nodes (TP8), with DSpark speculative decoding (k=5) and a 300K context, the Engram table kept in memory.
  12. 17 SeptemberNew LLM benchmarkDeepSeek-V4.1-Flash — 8× DGX SparkDeepSeek-V4.1-Flash (763B) on 8× DGX Spark: the same TP8 setup with a 1M-token context, the Engram table moved to disk.
  13. 16 SeptemberNew white paperDeepSeek-V4.1-Flash 4× DGX Spark DeploymentDeployment of DeepSeek-V4.1-Flash (763B MoE, FP8, DSpark k=5) on 4× NVIDIA DGX Spark (GB10) with tensor parallelism: vLLM build chain, 7 SM 12.1a patches, Engram-on-disk, benchmark results and B300 comparison.
  14. 16 SeptemberNew LLM benchmarkDeepSeek-V4.1-Flash — 4× DGX SparkDeepSeek-V4.1-Flash (763B) on 4× DGX Spark: FP8 weights served with vLLM across four nodes (TP4), with DSpark speculative decoding (k=5) and a 300K context.
  15. 16 SeptemberNew CV benchmarkPeopleNet INT8 — RTX 3060PeopleNet person detection (INT8, 960×544 input) on RTX 3060: frames per second per camera, measured from 1 to 16 simultaneous cameras.
  16. 14 SeptemberNew LLM benchmarkDeepSeek-V4.1-Flash — DGX B300DeepSeek-V4.1-Flash (763B) on DGX B300: FP8 weights served with vLLM on four GPUs (TP4), with DSpark speculative decoding (k=3) and a 1M-token context.
  17. 14 SeptemberNew LLM benchmarkTencent-Hy4-preview — DGX B300Tencent-Hy4-preview (780B) on DGX B300: BF16 weights served with vLLM on all eight GPUs (TP8), with MTP speculative decoding (k=3) and a 512K context.
  18. 14 SeptemberNew LLM benchmarkDeepSeek-V4-Flash-Vision-Exp — 2× DGX SparkDeepSeek-V4-Flash-Vision-Exp (305B) on 2× DGX Spark: FP8 weights served with vLLM across both nodes (TP2), with DSpark speculative decoding (k=3) and a 131K context.
  19. 4 SeptemberNew LLM benchmarkGLM-5.3 — 8× DGX SparkGLM-5.3 (753B) on 8× DGX Spark: NVFP4 weights served with vLLM across eight nodes (TP8), with the model's built-in MTP speculative decoding (k=4) and a 256K context.
  20. 4 SeptemberNew LLM benchmarkMuse-Glimmer-30B — RTX PRO 6000Muse-Glimmer-30B on a single RTX PRO 6000: BF16 weights served with vLLM, with DFlash speculative decoding through the model's native MTP (k=15).
  21. 1 SeptemberNew LLM benchmarkGLM-5.3 — DGX B300GLM-5.3 (753B) on DGX B300: FP8 weights served with SGLang on four GPUs (TP4), with a 1M-token context.

August 2026

  1. 3 AugustNew white paperNVIDIA DGX B300 vs GB300 NVL72 Cluster Architecture ComparisonTechnical comparison of two NVIDIA Blackwell Ultra architectures, DGX B300 and GB300 NVL72: system design, scaling approach, network fabric, power and cooling, and which workloads suit which platform.

July 2026

  1. 30 JulyNew white paperKimi K3 Inference Benchmark on DGX-B300Performance evaluation of Moonshot AI Kimi K3 (2.8T MoE, MXFP4) on NVIDIA DGX-B300 (8x Blackwell Ultra, TP=8): vLLM vs SGLang, direct vs DSpark speculative decoding, with SLO-driven capacity planning.
  2. 30 JulyNew white paperDGX Spark 3-Node AI Cluster Setup GuideRing (mesh) topology AI cluster setup with 3 NVIDIA DGX Spark nodes: management and compute networks, RoCEv2/RDMA, sparkrun configuration.
  3. 30 JulyNew white paperDGX Spark 2-Node AI Cluster Setup GuidePoint-to-point topology AI cluster setup with 2 NVIDIA DGX Spark nodes: management and compute networks, RoCEv2/RDMA, sparkrun configuration.
  4. 24 JulyNew white paperDGX Spark 8-Node AI Cluster Setup GuideSwitch-based AI cluster setup with 8 NVIDIA DGX Spark nodes over a MikroTik CRS804 with 200G breakout: management and compute networks, RoCEv2/RDMA, sparkrun and NAS.
  5. 24 JulyNew white paperDGX Spark 4-Node AI Cluster Setup GuideSwitch-based AI cluster setup with 4 NVIDIA DGX Spark nodes over a MikroTik CRS812: management and compute networks, RoCEv2/RDMA, sparkrun and NAS.
  6. 6 JulyNew white paperQwen3.6-27B DGX Spark Cluster ScalingMulti-node scaling study of Qwen3.6-27B-NVFP4 on 1x, 2x, and 4x NVIDIA DGX Spark (GB10): tensor parallelism over 200GbE, SLO-driven capacity planning, and TP-vs-replication deployment guidance.
  7. 3 JulyNew white paperQwen3.6-27B DGX Spark BenchmarkPerformance evaluation of the Qwen3.6-27B model on the NVIDIA DGX Spark (GB10) platform with FP8, FP8-MTP, AWQ-MTP, NVFP4, and NVFP4-MTP quantization variants.

June 2026

  1. 30 JuneNew white paperLocal LLM Usage GuideEnd-to-end decision guide for local LLM usage: hardware (NVIDIA Jetson, RTX PRO, DGX Spark, DGX/HGX), model selection, software stack and scenario mapping.