GEHIRNS

Training-free mixed-precision Qwen3.5-4B for 12 GB GPUs. With an 8-bit KV cache: 96% of BF16's score in 30% of the memory, 2.7x faster.

Comparison and retired builds: magi-beta-1 (alternative allocation) and magi-alpha-1/2 (retired fine-tunes of Qwen3.5-9B).