llama-modchip
Synchopate's llama-modchip v2.4.0 runs the 284B DeepSeek-V4-Flash model on a 64GB CPU-only system at 256K context, with 44.4 tok/s prefill and 9.1 to 14.7 tok/s decode.
The project combines a custom 1.71 bits per weight quantization format called Fractal-64 with SwiGLU-H2 compression and AVX-512 CPU kernels. This shrinks the DeepSeek-V4-Flash file to 56.56 GiB, compared with 76.87 GiB for standard UD-IQ1_S builds in the project's own comparison. Files need the project's patched runtime because they use custom tensor type 40, so stock llama.cpp, Ollama and LM Studio reject them.
- Date
- Sunday 8 November 2026
- Lab
- Synchopate
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| DeepSeek-V4-Flash total parameters | 284.33B (13.17B active) mixture-of-experts | company |
| Model file size | 56.56 GiB 1.71 bpw | company |
| Peak resident memory at 256K context | 57.61 GiB the news article quotes up to 58.9 GiB | company |
| Prefill speed, 284B | 44.36 tok/s 16-core Zen 5, CPU only | company |
| Decode speed, 284B | 9.1 to 14.7 tok/s range covers runs with and without Stockfish-PVS speculative decoding | company |
| Decode speed, 30B MoE models | 25.0 to 111.6 tok/s 3.75 bpw, depends on hardware and method | company |
All figures are project-reported with no independent validation. The code is Apache-2.0, and the first-release date is not confirmed, so the date is the day it was reported.
Sources
- github.com/synchopate/llama-modchip
- news.lavx.hu/article/synchopate-says-llama-modchip-runs-a-284b-model-on-a-64gb-cpu-system
This record was checked and corrected against its sources on 8 November 2026. How we check