# Can you run DeepSeek-V4-Flash on CPU with 64 GB RAM?

> Yes, with llama-modchip 2.4.0: a 56.56 GiB build of DeepSeek-V4-Flash runs on 64 GB of RAM and no GPU, at about 9 to 15 tokens a second per its author.

- Published: 2026-11-09
- Publisher: Rig Labs (https://riglabs.ai/)
- Canonical: https://riglabs.ai/blog/run-deepseek-v4-flash-on-cpu/
- Answers: run DeepSeek V4 Flash on CPU 64GB RAM

---

You can run DeepSeek-V4-Flash on a CPU with 64 GB of RAM using llama-modchip, an open-source project by GitHub user synchopate. Version 2.4.0, released on 2026-11-04, fits the 284B-parameter model into 57.61 GiB of memory with no GPU, according to [the project's README](https://github.com/synchopate/llama-modchip).

Every number below comes from the README, which lists the author's and community results.

## What is llama-modchip?

llama-modchip is a set of kernel patches, quantization tools and a runtime built on llama.cpp, released under Apache 2.0. Its README calls it "CPU-First Inference" for large Mixture-of-Experts models on 16 to 64 GB of RAM. It shrinks models with its own "Fractal-64" format, so the files need the project's patched binaries: stock llama.cpp, Ollama and LM Studio reject them with "unknown tensor type 40". [Release v2.4.0](https://github.com/synchopate/llama-modchip/releases/tag/v2.4.0) is dated 2026-11-04, and the project was discussed on [Hacker News](https://news.ycombinator.com/item?id=50411961) on 2026-11-08.

## How much memory does DeepSeek-V4-Flash need with llama-modchip?

DeepSeek-V4-Flash has 284B parameters, about 13B of them active per token. The README gives these figures:

| Build | File size | Fits in 64 GB? |
|---|---|---|
| Unpatched `UD-IQ1_S` | 76.87 GiB | No, "OOM / Swap on 64 GB" |
| llama-modchip v2.4.0 | 56.56 GiB | Yes, 57.61 GiB peak at 256K context |

The README says the build "Fits in 62.5 GiB usable Linux RAM" with no swap to disk.

## How fast does DeepSeek-V4-Flash run on a CPU?

On an AMD Ryzen 9 9955HX (Zen 5) with 64 GB of DDR5 and no GPU, the README reports:

- Reading a prompt: 43.16 to 44.36 tokens a second.
- Writing: 9.06 to 9.95 tokens a second by default, and 11.5 to 14.7 with its speculative "Stockfish-PVS" drafter.
- Faster modes use fewer experts per token (`LLAMA_MOE_TOP_K=5` or `4`), up to 13.0 to 16.0 tokens a second, with a small measured loss in quality.

## What hardware does llama-modchip support?

The README lists x86_64 CPUs with AVX-512 or AVX2, and ARM64 with NEON, SVE or Apple Metal. It runs on Linux and macOS through `./modchip chat`, and on Windows 10 and 11 through a hybrid GPU and CPU pack. Its sizing guide:

- 16 to 32 GB: Nemotron-3.5-30B-A3B, a 14.35 GiB build.
- 64 GB: DeepSeek-V4-Flash 284B.
- 128 GB and up: workstation class, such as Threadripper.

For the smaller Nemotron model, the README reports 25.02 tokens a second on the same Ryzen 9 9955HX, and 49.2 to 57.7 on an Apple M4 Max with 64 GB.

## How do you start it?

The README's quick start for Linux and macOS is to clone the repo and run its chat command:

```
git clone https://github.com/synchopate/llama-modchip.git
cd llama-modchip
./modchip chat
```

It also offers a prebuilt Linux x86_64 server and the model files on Hugging Face, as `Synchopate/DeepSeek-V4-Flash-0731-Fractal-64GB-CPU-GGUF`. The README's CPU check pipes a script from GitHub into Python. Read the script before you run it.

## What are the limits?

- The numbers are self-reported. The README shows no independent benchmark.
- The model files only run on the project's patched binaries, so your usual local tools cannot load them.
- 64 GB is the floor for the large model, with little room left for anything else on the machine.

Last checked 2026-11-09.

Sources:
- [llama-modchip on GitHub](https://github.com/synchopate/llama-modchip)
- [Release v2.4.0](https://github.com/synchopate/llama-modchip/releases/tag/v2.4.0)
- [Hacker News discussion](https://news.ycombinator.com/item?id=50411961)
