Serving Qwen3.8-Flash-Next at 262K context on 8 RTX 3090s hits 1,200 tok/s
QuixiAI · x · 2026-09-06
Eric Hartford (QuixiAI, creator of Dolphin and Samantha) details how he served Qwen3.8-Flash-Next — a 125B-parameter hybrid of Gated DeltaNet linear attention, Qwen Sparse Attention, n-gram embedding tables and a built-in speculative drafter — on eight five-year-old RTX 3090s with no FP8 or NVLink, at its native 262,144-token context with exact bf16 KV cache.
Measured benchmarks (1,000 in / 2,000 out per request):
- 1 concurrent: 140 tok/s, 14s median latency
- 8 concurrent: 590-600 aggregate tok/s
- 32 concurrent: peak 1,199.8 tok/s
A 512 GiB pinned-RAM second tier lets evicted conversations resume in under a second instead of re-prefilling. His first working config managed only 409.7 tok/s and used quantized KV that was silently corrupting answers; the final setup delivered a net 2.9x speedup while doubling context and removing the harmful quantization. Full configs are open-sourced.
More from Infra
- PhD students in some countries get B200s on demand, unlike their US peers — HildeKuehne · 2026-10-23
- Nvidia's Berlin GTC spotlights humanoid robots and Vera Rubin AI supercomputing — emmanuelvivier · 2026-10-23
- Lenovo adds 5 AI Library use cases, claims 1-week deployment and up to 70% lower token cost — shashib · 2026-10-23
- Your kernel doesn't need to beat cuBLAS on every shape, argues GPU dev — blelbach · 2026-10-23
- llama.cpp team launches Llama: one-click local AI on your Mac menu bar — Hugging Face · 2026-10-23
- China's daily token use hits 500T, up 4x as average price doubles to ¥13.46 — teortaxesTex · 2026-10-23