Breaking01 / AI
llama-modchip fits 284B DeepSeek model into 64GB RAM, no GPU
The file shrinks to 56.56GiB, but unpatched llama.cpp, Ollama and LM Studio cannot run it.

The developer of llama-modchip wrote that the tool ran a 284B-parameter DeepSeek model on 64GB of RAM without a GPU. The figures are for v2.4.0, released in November, and the test machine is a Ryzen 9 9955HX.
According to the README, the tool cuts the standard UD-IQ1_S build file by 20.31GiB (76.87GiB to 56.56GiB). It does so with patches to llama.cpp, an open-source program for running language models, and with conversion tools. At a 256K context window, runtime memory is 57.61GiB to 58.9GiB and swap is zero.
Speed. Prompt processing improved by 77.1% (25.05 to 44.36 tokens per second, developer-published figures), according to the release notes. Answer generation at default settings improved by 53.3% (5.91 to 9.06 tokens per second, developer-published figures).
The README states that the default settings compute all six paths of MoE (Mixture-of-Experts, a structure that selects only some expert neural networks). Logit cosine similarity, a measure of how closely outputs match, is 1.00000 in a developer-published figure. That value compares the same 56.56GiB file before and after the patch.
The 76.87GiB standard build exceeds the roughly 62.5GiB of usable RAM on a 64GB Linux system, the README says. It names page-cache thrashing or out-of-memory kills as the result.
Compatibility. The README confirms that unpatched llama.cpp, Ollama and LM Studio reject models in the Fractal-64 format with the error unknown tensor type 40. The dedicated ways to run them are the ./modchip command and a precompiled binary for Linux.
The developer's benchmark table does not include a comparison of answer quality between the 56.56GiB file and the 76.87GiB build.