I merged two Minimax-H3 Ref2VA finetunes using a custom spectral/geometric algorithm called PRISM:
Base A:
WarmBloodAban/Minimax-h3_Singularity(v2)Base B:
TenStrip/10Eros-Max(8-Step Turbo beta 6)
Why PRISM?
Merging fast turbo models into base models often introduces weight artifacts and signal degradation. PRISM handles this by:
Filtering diffuse noise delta using randomized SVD (rank 64).
Combining directional SLERP with log-space magnitude interpolation (Polar Geodesic).
Resolving sign conflicts and re-anchoring layer norms.
The turbo block distribution is weighted higher toward time/head blocks t_time: 0.75, t_head: 0.70) with an average turbo share of ~0.55.
Hugging Face Repo & Weights (fp16 + int8 + w6a8):
👉 https://huggingface.co/OliviaRossi/Minimax-h3_Singularity-x-10Eros-Max_PRISM
Feedback and test generations are very welcome!