<h2 id="i-merged-two-minimax-h3-ref2va-finetunes-using-a-custom-spectralgeometric-algorithm-called-prism:">I merged two Minimax-H3 Ref2VA finetunes using a custom spectral/geometric algorithm called PRISM:</h2><ul><li>Base A: <code>WarmBloodAban/Minimax-h3_Singularity</code> (v2)</li><li>Base B: <code>TenStrip/10Eros-Max</code> (8-Step Turbo beta 6)</li></ul><h3 id="why-prism">Why PRISM?</h3>Merging fast turbo models into base models often introduces weight artifacts and signal degradation. PRISM handles this by:<ul><li>Filtering diffuse noise delta using randomized SVD (rank 64).</li><li>Combining directional SLERP with log-space magnitude interpolation (Polar Geodesic).</li><li>Resolving sign conflicts and re-anchoring layer norms.</li></ul>The turbo block distribution is weighted higher toward time/head blocks <code>t_time: 0.75</code>, <code>t_head: 0.70</code>) with an average turbo share of ~0.55.Hugging Face Repo &amp; Weights (fp16 + int8 + w6a8):👉 <a target="_blank" rel="ugc" href="https://huggingface.co/OliviaRossi/Minimax-h3_Singularity-x-10Eros-Max_PRISM">https://huggingface.co/OliviaRossi/Minimax-h3_Singularity-x-10Eros-Max_PRISM</a>Feedback and test generations are very welcome!

Files with SHA256: a8023aa1770ffaf689e5141f655a0a4cd6688296f67fa01819d6f4b06a77f188