What's New
33 New Post-Training Algorithms
Onboarded and GPU-verified: CPO, AlphaPO, R-DPO, cDPO, SPO, DPNLL, MinorDPO, C2DPO, AlphaDPO, AOT, APO, BCO, BetaDPO, BPO, CalDPO, ChiPO, CPOSimPO, DiscoPOP, DPOP, DPOShift, DRDPO, EXO, FDPO, FocalPO, GPO, HDPO, Hinge, NCA, ODPO, RobustDPO, SamPO, SPPO, WPO.
Total algorithms: 51 (6 RL + 45 SL).
4 New Verified Models
- THUDM/glm-4-9b-chat-hf (9.4B, LoRA)
- google/gemma-3-4b-it (4.3B, full-tuning)
- moonshotai/Kimi-VL-A3B-Instruct (16.4B MoE, 2.8B active, LoRA)
- moonshotai/Kimi-VL-A3B-Thinking (16.4B MoE, 2.8B active, LoRA)
Total verified models: 38.
Other Changes
- Star prompt on first CLI run
- README appendix on rollout vs inference server design philosophy
- Model verification script (
scripts/verify_model.py) - 694 tests passing