GEHIRNS/magi-alpha-1
9B • Updated • 879
Comparison and retired builds: magi-beta-1 (alternative allocation) and magi-alpha-1/2 (retired fine-tunes of Qwen3.5-9B).
Note d1 SFT. Retired: chat degeneration (token salad, loops).
Note d2 SFT of Qwen3.5-9B. Retired: did not beat the untrained base at coding (LCB 48 vs 50/150).
Note Alternative allocation (attention k/v/output at Q6_K). 80.65%, ID 0.771. Use magi-1.0 instead.