Qwen3.6-35B-A3B (NVFP4)
qwen3.6-35b-a3b-nvfp4@spark-1
NVIDIA GB10 Grace Blackwell (DGX Spark)35B·3B activeNVFP4ctx 65,536measured 2026-07-28 06:24Z
Capped run. Reference tasks were limited to 100 samples. Not comparable to a full-dataset score.
Quality
ifevalreference
0.936100 samples2215sprompt_strict_acc 0.930prompt_strict_stderr 0.026prompt_loose_acc 0.920prompt_loose_stderr 0.027
per-sample outcomes · 93/100 passed
Serving performance
| conc | tok/s all | /stream | ttft p50 | ttft p95 | tpot |
|---|---|---|---|---|---|
| 1 | 27.6 | 27.6 | 321ms | 321ms | 35.2ms |
| 8 | 174.5 | 21.9 | 1929ms | 2069ms | 38.4ms |
| 32 | 320.9 | 10.4 | 6087ms | 9160ms | 78.3ms |
serving configuration
Results are only comparable between targets with identical flags. These are recorded from the live server, not from memory.
- enforce_eager
- true
- gpu_memory_utilization
- 0.55
- kv_cache_dtype
- fp8
- max_num_batched_tokens
- 4096
- moe_backend
- flashinfer_b12x
- reasoning_parser
- qwen3
- tool_call_parser
- qwen3_coder
- enable_auto_tool_choice
- true
notes
Live behind the `brain` and `115-model-local-moe` aliases. This is the model the bench exists to interrogate: it was deployed without measurement. FLAGS OF CONCERN: --enforce-eager disables CUDA graphs (costs throughput) and --gpu-memory-utilization 0.55 leaves ~45% of unified memory unused, capping KV cache and therefore concurrency. Both are worth an A/B once the baseline is recorded.