Synchopate says llama-modchip runs a 284B model on a 64GB CPU system
, updated First seen Updated Emerging, tapering
Meta · ChipsMETA$560.84▼4.8%
Reported by Hacker News and news.lavx.hu2 sources ↓
Key points
- The project says inference runs entirely in RAM, without NVMe swap
- It claims support for a 256K context
- Unpatched llama.cpp, Ollama and LM Studio cannot open the project’s custom tensor format
Sources
Details
The Hacker News post describes llama-modchip as software for running DeepSeek-V4-Flash on a CPU-only system with 64GB of RAM. The project also says the setup can run the model at a 256K context, but the post offers no independent measurement confirming those claims.
The project says users can start it with its command-line tool or use a precompiled binary offered through Hugging Face repositories and GitHub Release v2.4.0. Its models use a custom tensor type, which the post says is unsupported by unpatched versions of llama.cpp, Ollama and LM Studio.
Written from coverage as of
Images
- Hacker News
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh