# Open source claims to run a 284 billion-parameter DeepSeek model on a PC with 64GB memory without a GPU

- Link: https://github.com/synchopate/llama-modchip
- Discussion: https://startupxo.com/en/1255/
- Posted by startupxo at 2026-11-08T18:24:03.594Z
- 0 points, 1 comment, kind story, origin curated
- Written in ko, machine translated to en

## Summary

- llama-modchip, released by developer synchopate, is open source software (Apache-2.0) that compresses AI models further and adjusts CPU computation to let people run large AI models locally without a graphics card.
- According to the developer, DeepSeek V4 Flash shrinks from 76.87GiB to 56.56GiB, and peak memory use is about 57.6GiB even with a 256K context, so it fits on a 64GB PC. The original text says the existing 76.87GiB version stops due to insufficient memory on 64GB Linux.
- README figures show input processing at about 44tok/s and response generation at 9.06~10.74tok/s on an AMD Ryzen 9 9955HX (16 cores, 64GB DDR5). With an auxiliary draft technique, it reaches 11.5~14.7tok/s. These are the developer's own measurements.
- Nemotron 3.5 30B at the 30 billion scale was reportedly reduced to 14.35GiB and achieved 25~60tok/s on a PC with 16~32GB memory. The compressed models use a proprietary format and cannot be opened in standard llama.cpp, Ollama or LM Studio.

## Sources

- Hacker News: https://news.ycombinator.com/item?id=50411961 (237 points, 72 comments, checked 2026-11-08T18:22:59Z)

## Comments (1)

- **startupxo** (editorial), 0 points, 2026-11-08T18:24:03.594Z, #c528
  O: 64GB For developers who want to try a large model on a desktop without a GPU, the suggestion is to run the CPU check script in the README, then start by measuring the speed of the 14.35GiB 30B model.
  X: The compressed model uses the dedicated GGML_TYPE_Q2_0 format, so it cannot be opened in standard llama.cpp, Ollama, or LM Studio. It must be run using the provided dedicated binary or the modchip command.

JSON: https://startupxo.com/api/v1/item.json?id=1255&lang=en