H3-Tiny for ComfyUI

H3-Tiny transformers repackaged as single-file checkpoints for ComfyUI's native MiniMax H3 support. Part of H3-Tiny: structured pruning of MiniMax H3 for joint video + audio generation. Project page · Code · All checkpoints

File Parameters Workflow Converted from Steps
diffusion_models/h3_tiny_17b_ref2va_bf16.safetensors (34.2 GB) 17.05B Ref2VA H3-Tiny-17B-Pruned-Ref2VA 50
diffusion_models/h3_tiny_17b_fl2va_bf16.safetensors (34.2 GB) 17.05B FL2VA / T2VA H3-Tiny-17B-Pruned-FL2VA 50
diffusion_models/h3_tiny_14b_okd_ref2va_bf16.safetensors (28.1 GB) 13.98B Ref2VA H3-Tiny-14B-OKD-Ref2VA-v0.1 50

These are the same weights as the diffusers checkpoints, in ComfyUI's key layout. ComfyUI reads the pruned depth and FFN width from the tensors, so no custom node is needed.

Usage

  1. Put a file from diffusion_models/ in ComfyUI/models/diffusion_models/.
  2. Get the text encoder and VAEs from Comfy-Org/MiniMax-H3: text_encoders/qwen3vl_32b_minimax_h3_*.safetensors → models/text_encoders/, vae/minimax_h3_video_vae_fp16.safetensors and vae/minimax_h3_audio_vae_fp32.safetensors → models/vae/.
  3. Open ComfyUI's MiniMax H3 workflow (MiniMax H3 Image to Video for FL2VA / T2VA, MiniMax H3 Reference to Video for Ref2VA) and select the H3-Tiny file in Load Diffusion Model.
  4. Sample without the turbo LoRA: 50 steps, simple scheduler, guider without CFG (BasicGuider). We tested euler.

Use the ref2va file with the Reference to Video node and the fl2va file with the Image to Video node (with first and last frames for FL2VA, without for T2VA).

We tested these files on ComfyUI commit afcb984 (2026-11-21) at 1344×768 and 243 frames on one H200: the outputs match our diffusers pipeline. The few-step checkpoint (H3-Tiny-17B-FewStep-Ref2VA-v0.1) is not included, because ComfyUI's MiniMax H3 model has no input for its interval-time branch. The Comfy-Org turbo LoRAs were trained on the 33B model and are untested with H3-Tiny.

The files were made with scripts/convert_comfyui.py. It writes AdaLN in ComfyUI's curve form (adaln_t_table), which reproduces the diffusers AdaLN to about 1e-10 relative error.

License

These weights are Model Derivatives of MiniMax H3 and are distributed under the MiniMax H3 Community License Agreement; see NOTICE. Use of the weights must follow that agreement, including its territorial scope and acceptable-use policy.

H3-Tiny is an independent project and is not affiliated with or endorsed by MiniMax or Comfy Org.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including H3-Tiny/H3-Tiny-ComfyUI