Nov 8, 2026 · found Nov 9, 2026, 6:20 AM PT
llama-modchip claims DeepSeek V4 Flash at 9 to 14 tokens/s on a 64 GB CPU-only machine
Verdict: Vendor-reported · Model releases · Research
An open-source llama.cpp fork runs the 284B model in about 57 GiB by quantizing to 1.71 bits per weight. Speeds and quality are the author's own, and the format needs the project's own binary.
A llama.cpp fork claims DeepSeek’s 284B-parameter model runs without a GPU on a 64 GB desktop.
- Claim: DeepSeek-V4-Flash at 1.71 bits per weight uses about 57 GiB. On a Ryzen 9 9955HX the repo reports 9.1 to 9.95 tokens/s decode, and 11.5 to 14.7 with a speculative drafter. llama-modchip
- Evidence: The model has 284B total and 13B active parameters under an MIT license, so only the active experts do work per token. DeepSeek
- Caveats: The author ran the main tests and scored quality with a self-made 100-point test. The fastest settings prune experts and cost some quality, and the custom format is rejected by stock llama.cpp, Ollama and LM Studio. Nothing was run here. llama-modchip
- Why it matters: Single-digit speeds suit offline experiments, not interactive coding agents.
Verdict history
- Nov 9, 2026, 6:20 AM PT: Vendor-reported. Repo README and DeepSeek model card read; the date of the repo release is not shown. Nothing was run. Hacker News front page on 8 Nov (376 points). The repo's own memory figures differ (57.61 vs 58.9 GiB).