llama-modchip is an open-source set of patches and tools for llama.cpp that aims to run very large Mixture-of-Experts language models on an ordinary CPU. Its headline claim: DeepSeek-V4-Flash, a model with 284.33 billion parameters (13.17 billion active), running on a machine with 64 GB of RAM and no GPU. It is published by the GitHub account synchopate under the Apache 2.0 licence; the current version is v2.4.0. Posted on 8 November 2026, it reached 402 points and 180 comments on Hacker News; the repository itself is still small, with about 566 stars when we checked.
How it fits: the project compresses the model with its own "Fractal-64" format to 56.56 GiB (1.71 bits per weight), down from 76.87 GiB for the IQ2_XXS version it compares against, which runs out of memory on 64 GB Linux machines. It adds hand-tuned AVX-512 CPU kernels, a self-speculative drafter it calls "Stockfish-PVS", and saved KV-cache snapshots in /dev/shm that it says restore an agent's context in about 13 ms. A ./modchip command covers chat, a web UI, benchmarks, quantisation and launching Claude Code against the local model.
The numbers, all self-reported, come from one machine: an AMD Ryzen 9 9955HX with 64 GB of DDR5. Prompt processing runs at about 43–44 tokens per second; text generation at about 9–10 tokens per second in the default lossless mode and up to 11.5–14.7 with speculative decoding. Smaller supported models, such as Nemotron-3.5-30B (about 14 GiB, for 16–32 GB machines), are much faster. LavX News points out that no independent test has confirmed the results.
Limits: the model files use a custom tensor type, so unpatched llama.cpp, Ollama and LM Studio reject them with "unknown tensor type 40"; you need the project's binaries. Memory is tight: the README gives a peak of about 57.6–58.9 GiB at a 256K context, against roughly 59.6–62.5 GiB usable on a 64 GB system, leaving little room for anything else. The 284B kernels are written for AVX-512, and the README shows no 284B result on other CPUs. Faster "Top-5" and "Top-4" expert settings cost a little quality.
To try it on Linux or macOS: git clone https://github.com/synchopate/llama-modchip.git, then cd llama-modchip and ./modchip chat. The 284B model is on Hugging Face as Synchopate/DeepSeek-V4-Flash-0731-Fractal-64GB-CPU-GGUF, with a prebuilt Linux x86_64 server binary; Windows has a separate launcher.
