MiniMax-H3-FL2VA-Ref2VA-Hybrid-NVFP4

MiniMax H3 FL2VA x Ref2VA Hybrid NVFP4

This repository contains four experimental, single-file NVFP4 Hybrid checkpoints for MiniMax H3. They combine the FL2VA quality base with the Ref2VA reference-conditioning pathway using the tensor-selection strategy published by smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models.

All four checkpoints (b15-49, b20-49, b25-49, and b30-49) have completed manual ComfyUI validation with actual video generation. All four checkpoints also passed complete structural and per-tensor source verification. Generation examples and their settings are listed below.

Generation Examples

Browse the generation examples: two Ref2VA cases and two FL2VA cases, with input images, full prompts, shared settings, and output videos for each diffusion model. Within each case, only the diffusion model changes.

All four cases contain six outputs each, for a total of 24 videos. The Ref2VA cases use three and five reference images respectively; the FL2VA cases use a first frame and a first/last-frame pair respectively.

The current examples indicate that a single Hybrid checkpoint can handle both Ref2VA and FL2VA tasks, making it a practical default option for reducing model switching. This observation is limited to the four published cases, each using one fixed seed, and does not establish that Hybrid consistently matches or outperforms the original models. Differences across tasks remain to be explored through broader use.

Models

Block ranges are zero-based and inclusive. A lower start block applies the Ref2VA AdaLN projections to more of the Transformer, while a higher start block retains more FL2VA projections.

Merge Method

Each checkpoint uses minimax_h3_fl2va_pruned_nvfp4.safetensors as its base. For the selected inclusive block range blocks.<start>-49, these FP16 tensors are copied from minimax_h3_ref2va_pruned_nvfp4.safetensors:

blocks.<start>-49.adaln_proj.linear.weight
blocks.<start>-49.adaln_proj.linear.bias

All other tensors remain from FL2VA, including the FL2VA adaln_t_table, final-layer AdaLN and output heads, attention, MLP, normalization, embeddings, token refiners, and all 200 NVFP4 layers with their 800 packed weight, scale, and comfy_quant tensors.

No interpolation, dequantization, requantization, fine-tuning, or curve rebasing was performed. The b25-49 variant reproduces the smhfacct parameter-selection behavior in NVFP4 storage format; the other ranges apply the same verified selection rule with a configurable start block.

Sources

The original smhfacct b25-49 INT8 checkpoint was independently checked tensor by tensor: exactly 50 AdaLN tensors came from Ref2VA, while its remaining 882 tensors came from FL2VA.

ComfyUI Usage

Place one checkpoint under:

ComfyUI/models/diffusion_models/

Select it through the standard Load Diffusion Model / UNETLoader node and leave weight_dtype set to default. Use a recent ComfyUI build containing native MiniMax H3 and NVFP4 support.

Expected companion models:

text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
vae/minimax_h3_video_vae_fp16.safetensors
vae/minimax_h3_audio_vae_fp32.safetensors

All four checkpoints have been manually validated in ComfyUI and successfully generated videos. For b25-49, validation also explicitly covered both FL2VA first/last-frame workflows and Ref2VA reference-conditioning workflows, with video and audio output from both paths. Generation examples and their settings are listed below.

Validation Status

Completed on September 10, 2026 for all four checkpoints:

• input key, shape, dtype, metadata, and NVFP4 layer contracts passed
• each generated file reopened successfully with the standard safetensors reader
• all 1,132 output tensor hashes matched their intended source
• every selected AdaLN tensor exactly matched Ref2VA
• all non-selected tensors remained byte-identical to FL2VA
• all 800 NVFP4-related tensors remained byte-identical to FL2VA
• output file SHA256 hashes matched the values listed above
• upstream ComfyUI model-detection functions identified the format as MiniMax H3 with mixed-precision quantization

Runtime validation reported by the user for all four checkpoints:

• manual ComfyUI loading and sampling completed
• actual video generation completed successfully

Additional workflow and audio validation previously reported for b25-49:

• manual ComfyUI loading and sampling completed
• the Ref2VA reference-conditioning path produced video and audio output
• the FL2VA first/last-frame path produced video and audio output

The machine-readable report is included as verification.json.

Limitations

• These are experimental tensor-selection merges, not trained models.
• FL2VA and Ref2VA use different AdaLN timestep tables. These checkpoints deliberately keep the FL2VA table and directly transplant Ref2VA projections to match the smhfacct-compatible behavior. They do not use curve rebasing.
• NVFP4 and INT8 inference are not expected to produce pixel-identical trajectories.
• The four cases provide output videos generated with identical settings within each case, changing only the diffusion model.
• NVFP4 is intended primarily for NVIDIA Blackwell GPUs. Behavior and performance on emulated paths may differ.

License

These are modified Model Derivatives of MiniMax H3 and are distributed under the MiniMax H3 Community License Agreement. Review the geographic restrictions, use restrictions, and Acceptable Use Policy before downloading or using the models.

The license defines the Applicable Territory as worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States of America. Use, distribution, running, and display outside the Applicable Territory are not authorized by that agreement.

Modified checkpoint notice: this repository changes the source model by selecting the full AdaLN projection weight and bias from Ref2VA for the named Transformer block range and combining them with the FL2VA NVFP4 base. Each distributed filename includes hybrid and its selected block range.

Acknowledgements

• MiniMax for MiniMax H3
• Comfy-Org for native MiniMax H3 and quantized-model support
• lilcheaty for the source NVFP4 checkpoints
• smhfacct for the Hybrid block-selection strategy