A developer using the handle Synchopate claims in a GitHub project called llama-modchip that DeepSeek-V4-Flash runs on the CPU alone on a 64 GB machine. The project became popular on Hacker News on 8 November. The figures are the developer’s own measurements on a single machine with a 16-core AMD Ryzen 9 9955HX. WindowsForum found no independent replication of the result.

How the model fits in memory

DeepSeek-V4-Flash is a mixture-of-experts (MoE) model: it has 284.33 billion parameters, but uses 13.17 billion to compute each token. Each layer selects six of 256 experts plus one shared expert. The CPU therefore needs to read only a small fraction of the model weights per token.

llama-modchip is a set of modifications to llama.cpp. It quantises the model into a 56.56 GiB file averaging 1.71 bits per weight. Most experts are at around 1.56-bit precision.

The README names Unsloth’s UD-IQ1_S version as its starting point. That version is 76.87 GiB, or 82.5 GB, and does not fit in memory. Samcodeman.com and WindowsForum described the starting version as an IQ2_XXS quantisation. Samcodeman.com reported that a Hacker News commenter pointed out that Unsloth’s IQ2_XXS file is 91 GB.

With a 256,000-token context window, memory use is 57.61–58.9 GiB. The README says about 62.5 GiB is available on a 64 GB Linux machine. WindowsForum put the available memory at about 59.6 GiB, which leaves 0.7–2 GiB for everything else.

Speed and quality

Prompt processing sped up from 25.0 to 44.36 tokens per second, the README says. With default settings, the model generates responses at 9.06–9.95 tokens per second, compared with 5.91 for the unmodified starting version. Samcodeman.com notes that the README also records the same starting version running out of memory or paging on a 64 GB machine.

With default settings, speculative decoding gives the fastest result, 11.5–14.7 tokens per second. A 1.64 MB Stockfish-PVS drafter proposes tokens, and the large model checks them. The gain depends on the share of proposals accepted, so speed varies by task.

In agentic use, prompt processing is the slowest step. Samcodeman.com calculated that processing a new 10,000-token prompt takes about 225 seconds. The program restores previously processed context in 13.1–32.6 milliseconds, the README says. This helps only if the beginning of the prompt stays the same.

The README calls the default setting lossless. However, it compares against the same quantised model without the project’s modifications, not against the original model. For the 284B model, the README reports neither a comparison with the full-precision model nor task tests. Samcodeman.com reported that a Hacker News commenter said even a 97 GB quantisation differs clearly from the full-precision model.

If each layer uses four experts instead of six and the computation takes other shortcuts, generation speeds up to a maximum of 10.74 tokens per second. The output then differs slightly from the default setting.

Getting started

The files use a custom tensor type, so standard llama.cpp, Ollama and LM Studio cannot open them. The project is licensed under Apache 2.0. Users download the model files and a prebuilt Linux x86_64 server from Hugging Face, and the instructions are in the repository’s README. The speed-ups for the 284B model target AVX-512BW instructions. For 16–32 GB machines, the README offers the 14.35 GiB Nemotron-3.5-30B-A3B model.