POCKET-Darwin-180B: 180B model runs on a laptop without a GPU

VIDRAFT released POCKET-Darwin-180B, a compressed 4-bit build of Darwin-180B-RSI, the model topping seven Hugging Face leaderboards. The 111 GB build, down from 360 GB, runs on a laptop with 8 GB VRAM and 32 GB RAM (4.17 tokens/s) or a CPU-only server (18.4–21.0 tokens/s) with identical MMLU-Pro accuracy of 87.65%.

• Compressed model is 111 GB (4-bit GGUF) versus 360 GB in BF16
• CPU-only server with 16 threads hits 18.4–21.0 tokens/s at 78.8 GB peak memory
• Laptop with RTX 5060 Laptop and 32 GB RAM reaches 4.17 tokens/s
• MMLU-Pro accuracy unchanged: 87.65% versus 87.65%