LLM Inference Benchmark Explorer
DeepSeek-V4.1-Flash on 4× DGX Spark — FP8, vLLM, TP4, speculative decoding k=5 inference benchmark
Open this configuration in the LLM Inference Benchmark Explorer
| Model | DeepSeek-V4.1-Flash |
|---|---|
| Parameters | 763B |
| Intelligence Index | 39.5 |
| Agentic Index | — |
| Device | 4× DGX Spark |
| Quantization | FP8 |
| Max C | 2 |
| Chat Capacity | 8 |
| Agentic Capacity | 3 |
| TP | 4 |
| DP | — |
| PP | — |
| Engine | vLLM |
| Speculative Decoding | Yes |
| C | TTFT (ms) | TPS (tok/s) | Status |
|---|---|---|---|
| 1 | 272 | 29.48 | PASS |
| 2 | 396 | 21.32 | PASS |
| 4 | 577 | 13.09 | FAIL |
| 8 | 806 | 8.79 | FAIL |
In OpenZeka's measurement, DeepSeek-V4.1-Flash (763B parameters), served in FP8 format with vLLM and speculative decoding on 4× DGX Spark (TP=4), reached a generation speed of 29.5 tok/s per request and a time to first token (TTFT) of 272 ms with a single request.
Considering the speed targets alone (the KV cache limit was not calculated for this configuration), the estimated capacity is 8 users for chat use and 3 for agentic use. The realistic capacity will likely fall between these two values. In scenarios dominated by coding, tool use, long workflows and multi-agent use, capacity approaches the agentic estimate; where shorter interactions, standard conversations and lighter tasks dominate, it approaches the chat estimate.
Notes: DSpark k=5, 300K context. Model: deepseek-ai/DeepSeek-V4.1-Flash. vLLM vllm-dsv41:latest.
Intelligence Index and Agentic Index values are published by Artificial Analysis and are reproduced here with attribution.
In OpenZeka's measurement, DeepSeek-V4.1-Flash (763B parameters), served in FP8 format with vLLM and speculative decoding on 4× DGX Spark (TP=4), reached a generation speed of 29.5 tok/s per request and a time to first token (TTFT) of 272 ms with a single request.
Considering the speed targets alone (the KV cache limit was not calculated for this configuration), the estimated capacity is 8 users for chat use and 3 for agentic use. The realistic capacity will likely fall between these two values. In scenarios dominated by coding, tool use, long workflows and multi-agent use, capacity approaches the agentic estimate; where shorter interactions, standard conversations and lighter tasks dominate, it approaches the chat estimate.