Testing Qwen3.6-27 Q8: Nearly 100K Context on 32GB VRAM
BitGreen1270 · reddit · 2026-07-05
The author shares benchmark parameters for running the Qwen3.6-27B Q8 quantized model逼近 100K context on an Ubuntu headless server (5090+64GB RAM) using llama-server. Subjectively, Q8 performs better than Q6/Q5 at around 60K context. VRAM is further optimized using configurations like KV Q80 and -b/-ub. The author notes this isn't rigorous research, just experience from a personal vibe coding project, and the VRAM is near its limit.
More from Infra
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Lam Research: weekly token usage up ~250x in 18 months — Beth_Kindig · 2026-10-27
- AWS shows how to build enterprise agentic retrieval on Bedrock with CDK — AWS ML Blog · 2026-10-27
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27