Llama-modchip claims a 64GB PC can operate a 284B DeepSeek model

The project llama-modchip claims that a PC with 64 to 128 GB of memory can operate a 284 billion parameter model without a graphics processor. The model is DeepSeek-V4-Flash, and the memory type is DDR5. The project claims that it works with 256K context, the text that the model reads together. It claims this removes multi-GPU hardware that costs $30,000 or more.

How it works

The tool checks the processor cores, the free memory and the graphics memory. It downloads software that operates models on a local PC. It divides the model between the processor and the graphics processor. Then it opens a chat page on the PC.

The numbers

The text gives no test results. Users send the score of a benchmark command to a leaderboard that all can see. A second model is the 30 billion parameter Nemotron-3.5-30B-A3B model.

What is not known

The text gives no speed and no quality result. It tells of guides with proofs, but it shows none. The project wants sponsors: hardware makers for test units, cloud teams for credits, and companies that pay for model conversion.

Insight: Speed in tokens a second decides if a local model works

A laptop test found that a 120 billion parameter model made 65.4 tokens a second on Windows and 13.9 on Linux. The project gives no such number. Software changes speed as much as hardware does. The project wants hardware makers to sponsor units for its leaderboard, and it seems to want to be the scoreboard that all use. Measure tokens a second on your hardware.