One RTX 3090 Ti runs a 552B DeepSeek MoE at 35 tok/s — full local LLM guide
uniVocity · reddit · 2026-11-02
A detailed Reddit guide runs DeepSeek-V4.1-Flash-1017 (552B MoE, 39B active) at 35 tok/s and Qwen3.8-Flash-Next at 138 tok/s on a single RTX 3090 Ti + 128GB RAM. Key trick: the open-source gpu-unlock kernel module removes artificial FP16/BF16 throttles on consumer Ampere cards, with full install commands, NVIDIA driver options, quantization choices, and Gentoo dependency setup.
More from Infra
- OpenAI may rebrand its data centers as 'AI compute campuses' — TechCrunch AI · 2026-11-02
- $300B Louisiana project pipeline squeezes steel supply, H-beam lead times hit 9 months — MatthewChang · 2026-11-02
- Benchmark: Speedy inference engine crushes Avalon — 1.6x faster reads on Qwen3 long context — MotorNetwork380 · 2026-11-02
- vLLM serves DeepSeek-V4.1-Flash at 954K tokens/s per GPU on one GB300 NVL72 rack — bookwormengr · 2026-11-02
- Prime Mega MoE extends DeepSeek's kernel for training, up to 34% faster end-to-end MoE training — cfregly · 2026-11-02
- 2.3bpw quant update: 27B model on 12GB cards ties Qwen3.6 BF16 on LiveCode Bench at 84.3% — Brief-Tap-6616 · 2026-11-02