Bits fitted to each model by measurement. magi-1.2-2B: loop-DPO + 2B map, density 1.305. magi-1.2-4B: 4B map, density 0.968.
AI & ML interests
None defined yet.
Recent Activity
Training-free pruned and quantized Jeffan-4B-2010 decision model. At 8K context: 98% of its score in 56% of the GPU memory, 1.4x faster.
Comparison and retired builds: magi-beta-1 (alternative allocation) and magi-alpha-1/2 (retired fine-tunes of Qwen3.5-9B).
Bits fitted to each model by measurement. magi-1.2-2B: loop-DPO + 2B map, density 1.305. magi-1.2-4B: 4B map, density 0.968.
The magi-1.0 recipe plus a LoRA that stops loops and finishes thinking. magi-1.1-4B: 81.7%, density 0.939. magi-1.1-2B: density 1.149.
Training-free pruned and quantized Jeffan-4B-2010 decision model. At 8K context: 98% of its score in 56% of the GPU memory, 1.4x faster.
Training-free mixed-precision Qwen3.5 for 12 GB GPUs: magi-1.0-4B (96% of BF16's score in 30% of the memory) and magi-1.0-2B.
Comparison and retired builds: magi-beta-1 (alternative allocation) and magi-alpha-1/2 (retired fine-tunes of Qwen3.5-9B).