Openzeka Technical White Papers
In-depth technical analysis on artificial intelligence, high-performance computing, and embedded systems. From NVIDIA DGX/HGX data center solutions to Jetson-based edge inference, from DGX Spark mini supercomputers to the Cordatus AI real-time video analytics platform.
Latest White Papers
What 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.
October 2026 Technical GuideNetwork 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.
October 2026 Decision GuideLLM 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.
October 2026 Device AdvisorDevice AdvisorDescribe your cameras and pick a vision-language model, and this page names the smallest device we have measured that keeps up — then plays the stream back as a live timeline, so you can see how often each camera is answered and how many of its frames the model ever sees.
October 2026 VLM BenchmarkVLM Inference Benchmark ExplorerExplore vision-language model inference benchmarks on Jetson AGX Thor, DGX Spark, RTX PRO 6000 Blackwell, Jetson AGX Orin 32GB and Jetson Orin NX 16GB. Pick the image size, set how long a camera may wait for its answer, and see how many cameras each configuration keeps up with.
October 2026 Technical GuideThe 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.
October 2026 Technical GuideEngram 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.
September 2026 LLM DeploymentDeepSeek-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.
September 2026 CV BenchmarkCV Inference Benchmark ExplorerExplore OpenZeka's computer-vision inference benchmarks on NVIDIA GPUs and Jetson devices. Pick a GPU and a detection model to see the frame rate it sustains, how it falls as cameras are added, and how many cameras it carries at your target FPS.
September 2026 LLM DeploymentDeepSeek-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.
September 2026 LLM BenchmarkLLM Inference Benchmark ExplorerExplore OpenZeka's LLM inference benchmarks on NVIDIA DGX Spark, DGX B300, RTX PRO 6000 Blackwell and Jetson Thor. Filter by model, parameter count, device, quantization and concurrency, and set your own performance targets.
August 2026 Architecture ComparisonNVIDIA 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.
August 2026 LLM BenchmarkKimi 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.
July 2026 Cluster SetupDGX 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.
July 2026 Cluster SetupDGX 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.
July 2026 Cluster SetupDGX 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.
July 2026 Cluster SetupDGX 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.
July 2026 LLM ScalingQwen3.6-27B 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.
July 2026 LLM BenchmarkQwen3.6-27B 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.
July 2026 Decision GuideLocal 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.
June 2026Latest changes
- 21 October 2026New 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.
- 21 October 2026New 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.
- 21 October 2026New 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.
- 21 October 2026New 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.
- 21 October 2026New 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.
Areas of Expertise
Embedded & Edge AI
Real-time inference, robotics, and physical AI on the NVIDIA Jetson module family (Orin, Thor, Xavier).
Data Center
Scalable AI training and inference infrastructure with DGX, HGX, and PCIe-based server systems.
Mini Supercomputer
DGX Spark family with GB10 architecture — compact, powerful, and scalable personal AI workstations.