Radar ·
Llama.cpp fork reports a 284B model running on CPU with 64GB of RAM
A modified llama.cpp fork, called llama-modchip, is reported to run a 284B total-parameter model on 64GB of DDR5 RAM without a GPU, using low-bit quantization and custom AVX2 and AVX-512 kernels.
The report describes llama-modchip, a modified llama.cpp fork, which reportedly runs a model with 284B total parameters, about 13B active, on 64GB of DDR5 RAM and no GPU. The gains are attributed to low-bit quantization and custom AVX2 and AVX-512 kernels. The source presents CPU-only inference of a large model as a performance demonstration.
The figures are reported by the developer and were not verified in this note. The source gives no full methodology, latency, output quality or hardware setup beyond the RAM cited. To check, consult the fork's original repository and documentation, look for independent benchmarks, and reproduce the test in your own environment before drawing conclusions. This note does not indicate that Rota Nacional runs this model or uses this fork.