Merging 2 AI models

Minimax-h3_Singularity-x-10Eros-Max_PRISM

Merge of two Minimax-H3 Ref2VA finetunes with a custom algorithm (PRISM):

  • A: WarmBloodAban/Minimax-h3_Singularity / Minimax-h3_Singularity_ref2va_v2_bf16.safetensors
  • B: TenStrip/10Eros-Max / 10Eros_Max_h3_8stepTURBO_beta6_bf16.safetensors (8-step turbo)

Weights: Minimax-h3_Singularity_x_10Eros-Max_PRISM_fp16.safetensors (fp16). Licence follows the source models - check them before use.

Method

  1. Spectral denoise - D=B-A split by randomised SVD into structured part L (rank<=64) and diffuse remainder R; B' = A + L + gamma*R with adaptive gamma in [0.92, 1].
  2. Polar geodesic - per output row, direction by SLERP, magnitude by log-domain (geometric) interpolation.
  3. Sign-conflict mediation - salient opposite-sign elements are pulled toward the weighted winner.
  4. Magnitude re-anchor - rows rescaled to the geometric target norm.

Share of B (turbo): {'t_early': 0.45, 't_late': 0.65, 't_default': 0.55, 't_time': 0.75, 't_head': 0.7, 't_ref': 0.45} (blocks ramp smoothly early->late; mean ~0.55). See merge_report.json for diagnostics (covers the tensors merged in the final run if the job was resumed).

Downloads last month
11,413
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for OliviaRossi/Minimax-h3_Singularity-x-10Eros-Max_PRISM