Llama-modchip runs 284B DeepSeek on 64GB RAM without GPU
The gist
GitHub developer synchopate released llama-modchip, running a 284.33B DeepSeek model on 64GB of CPU memory with no GPU.
The benchmarks are the project's own, and the compressed format needs its patched runtime.
How it unfolded
- Nov 6, 2026David Hendrickson (@TeksEdge) posts on X that synchopate has published a modified llama.cpp stack running DeepSeek-V4-Flash entirely in 64GB DDR5 memory with zero GPUs, citing developer-reported figures including a 26.4% reduction in model storage.
- Nov 8, 2026LavX News reports the project lists a 56.56 GiB compressed model file and up to 58.9 GiB resident memory at 256K context, with 9.06–10.74 tok/s generation without speculative decoding on an AMD Ryzen 9 9955HX with 64GB DDR5.
- Nov 8, 2026Tech Bytes reports Synchopate launched llama-modchip v2.4.0, running the 284.33B model on 64GB CPU systems at up to 14.7 tokens per second.
Key terms
- mixture-of-experts
- A machine learning technique using multiple expert networks to divide a problem space into regions.
- quantization
- Mapping input values from a large set to a smaller finite set of discrete values, as in rounding or lossy compression.
Source: Wikipedia