Qwen3.8-Flash-Next
Throughput
Decode tok/s after a fixed context fill (PBM) — same agent-style workload as bench v2.
Published evals
Vendor-card figures for qwen/qwen3.8-flash-next — not this pack. Not measured on this Spark.
- Terminal-Bench 2.1
- 82.9% NVIDIA NVFP4 model card ↗ 2026-08-31 NVIDIA NVFP4 column; temp=1.0, top_p=0.95, reasoning_effort=xhigh. FP8 baseline on the same card is 83.3. Not the Mia pack, not measured on Sparky.
- SWE-bench Pro
- 62.5% Qwen model card ↗ 2026-08-26 Claude Code harness; refined SWE-bench Pro set (vendor). DeepSWE 1.1 is 58.7 on the same card (not a site column).
- LiveCodeBench
- 91.9% Qwen model card ↗ 2026-08-26 LiveCodeBench v6 (vendor).
- GPQA Diamond
- 91.7% Qwen model card ↗ 2026-08-26
Independent board on Artificial Analysis ↗ — we do not copy their scores.
Recipe
- Profile
- mia-ailab-qwen3.8-flash-next-vllm
- Engine
- flashnext
- Context
- 256k · fp8 KV
- Served as
- qwen3.8-flash-next
- Draft
- MTP · n3
Run on your Spark
spark inference up mia-ailab-qwen3.8-flash-next-vllm
Why we run it
Qwen3.8-Flash-Next 125B-A6B VLM on one Spark via MiaAI NVFP4 (~99 GiB) plus a first-boot PLE mmap table (~27 GiB). Official NVIDIA NVFP4 ~135 GB does not fit. Served by spark engine flashnext (vllm/vllm-openai:qwen38-flash-next), not eugr. Companion A/B to golden 27B SGLang DFlash2. Site row is native PBM 36.1/21.6/19.6; recipe stays testing until bench v2.
Bench notes
PBM 4k @ 36.1 / 50k @ 21.6 / 100k @ 19.6 tok/s — perfbench-metrics — native 262k MTP3 vision on — profile=mia-ailab-qwen3.8-flash-next-vllm — YaRN 512k companion 36.3/22.0/18.1 plus 200k=31.3 / 300k=12.7