Community-Built PRISM Merges MiniMax-H3 Quality With Eight-Step Speed
A community-published merge for MiniMax-H3's Ref2VA video model uses SVD denoising and geodesic interpolation to fuse a quality finetune with an 8-step turbo variant.
- New Hugging Face merge Minimax-h3_Singularity-x-10Eros-Max_PRISM combines two Ref2VA finetunes of MiniMax-H3.
- Custom PRISM algorithm uses randomised SVD to split structural from diffuse weight deltas.
- Direction merged via SLERP, magnitude via log-domain geometric interpolation, with sign-conflict mediation.
- Per-block mixing ramps turbo contribution from 0.45 early to 0.75 on time embeddings.
- Built on MiniMax-H3, an open-weight 24fps video plus stereo audio generator.
- Already 2,400+ downloads, fp16 safetensors, no benchmarks or sample outputs published.
A community-built Hugging Face checkpoint combines a quality-oriented MiniMax-H3 reference model with an eight-step turbo variant. The PRISM model card describes a four-stage algorithm designed to preserve reference fidelity while reducing the denoising-step budget. Hugging Face showed more than 2,400 downloads at publication.
MiniMax-H3 is an open-weight multimodal model that generates 24 FPS video with native stereo audio from text, image, video, and audio references. Its FL2VA family conditions generation on first and last frames. Ref2VA injects encoded reference tokens into every denoising step, helping carry identity, style, motion, and voice into new scenes. A single Ref2VA request can include up to nine images, three videos, and three audio clips.
One file, two H3 lineages
PRISM draws its reference behavior and turbo schedule from two Ref2VA fine-tunes:
| Artifact | Model | Role | Precision |
|---|---|---|---|
| Model A | Singularity | Quality-oriented Ref2VA fine-tune | bf16 |
| Model B | 10Eros-Max | 10Eros_Max_h3_8stepTURBO_beta6_bf16 |
bf16 |
| Output | PRISM | Single merged safetensors checkpoint |
fp16 |
Both bf16 and fp16 are 16-bit floating-point formats with different range and precision trade-offs, so loaders should preserve the published output type. The card lists the license as inherited from the source models. Redistribution and commercial-use rights depend on those parent terms.
Inside the four-stage merge
The model card describes PRISM as a sequence of spectral filtering, row-wise interpolation, sign mediation, and norm correction:
- Spectral denoising. For eligible weight matrices, PRISM computes the delta
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
