Running Qwen3.8 at 170K Context on a Single 96GB GPU
UltrMgns · reddit · 2026-08-30
A user shared a practical deployment of Qwen3.8-Flash-Next on a single 96GB GPU, achieving a 170K context window and 110 tokens/sec.
Configuration & Optimizations:
- Used an INT4 quantized version (32GB) via memory mapping.
- Deployed with vLLM, enabling Speculative Decoding (MTP), Prefix Caching, and Chunked Prefill.
- Modified the model.safetensors.index. to drop unnecessary ple-bf16 weights.
- Set --max-model-len 173400, utilizing 89GB VRAM on a single Pro 6000.
Performance:
- The model successfully coded and played complex HTML games in a headless Ubuntu environment, continuously improving its strategy.
- Achieved 87% MTP acceptance rate on code/JSON and >90% prefix cache hit rate on long tasks.
Related event: Qwen3.8 Runs 170K Context on Single 96GB GPU(2 posts)→
More from Infra
- Grok claims 816B MoE model runs locally on dual M5 Ultra Mac Studios via MLX — grok · 2026-10-16
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16
- Google Raises 2026 Capital Spending to Up to $205 Billion — emmanuelvivier · 2026-10-16