# Llama\-modchip: 284B DeepSeek on 64GB RAM, no GPU

- HackerByte page: <https://hackerbyte.net/en/stories/a-project-enabling-the-operation-of-the-284-billion-parameter-deepseek-llm-model-50411961>
- Original source: <https://github.com/synchopate/llama-modchip>
- Hacker News discussion: <https://news.ycombinator.com/item?id=50411961>
- Submitted to Hacker News: 2026-11-08T10:22:04.000Z

A project enabling the operation of the 284 billion parameter DeepSeek LLM model on standard 64GB RAM, without requiring a GPU.

## AI-generated summary

Llama\-modchip is a project that enables running large language models, specifically the 284B DeepSeek\-V4\-Flash and 30B MoE models, on systems with only 64GB of RAM and no dedicated GPU. It achieves this through a combination of kernel patches, advanced quantization techniques like Fractal\-64, and optimized runtime tools.

## Key points

- Enables running 284B DeepSeek\-V4\-Flash on 64GB RAM using a 56.56 GiB quantized model.
- Achieves 44.4 tok/s prefill and 9.1\-14.7 tok/s for the 284B model without a GPU.
- Supports 30B MoE models with significantly higher throughput \(25\-111.6 tok/s\) on less RAM.
- Utilizes advanced quantization \(e.g., Fractal\-64 with 1.71 bpw for DeepSeek\) and CPU\-specific kernel optimizations \(AVX\-512, AMX, NEON\).
- Offers features like self\-speculative drafting and rapid context restore for agent\-like performance.