GoKawiilOpen-source llama-modchip runs 284B-parameter DeepSeek model on 64GB CPU-only machine
A project called llama-modchip, released on GitHub, claims to compress the 284.33B-parameter DeepSeek-V4-Flash mixture-of-experts model from roughly 77GB down to about 57-59GB of memory, letting it run on a 64GB CPU-only Linux machine without a GPU. The developers report prefill speeds near 44 tokens per second and CPU decode speeds of roughly 9-15 tokens per second, achieved through custom quantization, cache-tiling and pruning techniques, plus a separate mode running a 30B Nemotron model at much higher speeds on CPU hardware.
GoKawiil's interpretation of the reporting above, not reported fact.
If the claimed compression and speed figures hold up under independent testing, it would make frontier-scale mixture-of-experts models usable on commodity CPU hardware rather than requiring expensive GPU clusters, which could broaden access for hobbyists and smaller organizations. The project also uses nonstandard tensor formats incompatible with mainstream tools like llama.cpp or Ollama, suggesting adoption may depend on the maintainers' own forked toolchain rather than the broader ecosystem.
- llama-modchip claims to run a 284B-parameter DeepSeek model in roughly 57-59GB of RAM without a GPU
- Reported CPU-only decode speeds range from about 9 to 15 tokens per second depending on configuration
- The project also offers a 30B Nemotron model mode with reported speeds up to 111 tokens per second on Threadripper CPUs
Corsair Vengeance 64GB DDR5 RAM Kit — This article is all about squeezing massive 284B-parameter MoE models into a 64GB RAM box without a GPU, which means your system memory is the real bottleneck. A solid 64GB DDR5 kit like this gives you the headroom to actually try llama-modchip style CPU inference workloads yourself. Great for anyone experimenting with large local LLMs on consumer hardware.
See Corsair Vengeance 64GB DDR5 RAM Kit on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: github.com, 2026-11-08
Published there as: “Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.