VIDRAFT releases POCKET-Darwin-180B for CPU-only machines

On October 3, 2026, DEV Community's AI section reported that VIDRAFT had released POCKET-Darwin-180B, a 4-bit GGUF-quantized build of the company's Darwin-180B-RSI frontier model. According to the report, the release is compatible with llama.cpp and runs on consumer hardware, including CPU-only laptops and mini PCs, rather than the enterprise GPU clusters usually associated with models of this size. The article frames the news around running a 180B-parameter LLM on a laptop without a GPU.

The compression is attributed to two techniques working together: sparse MoE routing and graft quantization. Together they shrink the original BF16 model from 360 GB to 111 GB, packaged as only 4 files. The parameter count is 180B.

Beyond those points the material is thin. The report does not list minimum hardware requirements, expected inference throughput, memory needs, or any benchmark results for the quantized build. It also does not indicate whether POCKET-Darwin-180B arrived before or after the original Darwin-180B-RSI, and it gives no download or licensing details. Only this single report dated October 3, 2026 is available here, so the account of the announcement rests on it alone.