GEHIRNS/magi-1.0-4B
Text Generation • 4B • Updated • 1.48k • 5
MAGI 1.0
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.
Note Recommended. 1.80 GiB GGUF, ID 0.893 with an 8-bit KV cache (3.0x BF16), 181 tok/s on an RTX 5070.
GEHIRNS/magi-1.0-2B
Text Generation • 2B • Updated • 537Note The magi-1.0 recipe on Qwen3.5-2B, no training. 0.83 GiB GGUF, density 1.071 at 32K with a q4_0 KV cache (magi-1.0: 0.903 at the same settings), 332 tok/s on an RTX 5070.