Ninfer6000 benchmarked on Qwen3.8-Flash-Next: 1.59x decode speed over vLLM
dboybaker · reddit · 2026-10-16
A blogger benchmarked three inference engines running Qwen3.8-Flash-Next on an RTX Pro 6000 Blackwell: Ninfer6000 (a Blackwell-tuned Ninfer fork), vLLM (NVFP4), and llama.cpp (Q4KXL).
Speed:
- Decode: Ninfer6000 averages 238 tok/s — 1.59x vLLM (150) and 2.21x llama.cpp (108); the gap vs llama.cpp widens to 3.7x at 114k-token prompts
- Prefill: vLLM still leads (16k tok/s), Ninfer6000 10k, llama.cpp 3k
- The author didn't reproduce the 400 tok/s from an earlier post but got close in idealized conditions
Accuracy: BFCL v4, IFEval, MMLU-Pro, HumanEval+, MBPP+, Arena-Hard and RULER are all within noise across engines; only tool calling (76.67 vs 74) favors Ninfer6000 beyond noise. The author did hit notable tool-calling formatting issues in real-world use.
More from Infra
- Vercel Sandbox Drives go GA: attach TBs of data in sub-second, fork instantly — cramforce · 2026-10-16
- NVIDIA launches RTX Spark for local AI on Windows PCs — ostrisai · 2026-10-16
- SkyRL ships Day-0 RL support for Beam with +62% LoRA decode throughput in vLLM — vllm_project · 2026-10-16
- DFlash2 delivers up to 3.03x decode-only speedup for Beam across five benchmarks — vllm_project · 2026-10-16
- vLLM's Beam serving kernels cut GB200 forward time from 8.15 to 6.47 ms at TP8 batch 128 — vllm_project · 2026-10-16
- Reflection's 501B MoE Beam hits 463 tok/s on 8x GB200 with NVFP4 at day 0 — zhyncs42 · 2026-10-16