MiniMax H3 Veda Acceleration | 3x Fast Ref-to-Video
Veda Sparse Attention (MiniMax H3)
minimax_h3_r2va_veda_preview_fp8.safetensors
Replace with minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors
ComfyUI model file · Sparse-attention predictor
Veda sparse-attention predictor for MiniMax H3 · Up to 3.1× faster MiniMax H3 sampling
This is the Veda predictor for MiniMax H3. It learns which 10% of attention matters and lets the model skip the rest, for up to 3.1× faster end-to-end generation. Put it in ComfyUI/models/veda and select it in the Veda Sparse Attention node. If you searched for h3_t2va_deda_8nfe_600step_preview_fp8, this is the file.
There is no file named h3_t2va_deda…; “deda” is a common misspelling of Veda. The real filename starts with minimax_h3_t2va_veda.
minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors
The short version
Attention is most of the cost of video diffusion, but only a small share of it changes the result. Veda trains a light predictor, distilled from the full model, to pick the most important 10% of attention so H3 can skip the rest. The authors report up to 6.8× faster attention and 3.1× faster end to end, with quality on par with full attention.
It is not a LoRA: model weights and style stay the same, so it combines with any LoRA or fine-tuned H3. “T2VA” and “8NFE 600step” describe how this preview checkpoint was trained, not where it can be used.
Searches for h3_t2va_deda_8nfe_600step_preview_fp8.safetensors land here because no file by that name exists; it is a misspelling of this one.
Install
ComfyUI/ └─ models/ ├─ … └─ veda/ └─ minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors
$ cd ComfyUI/models/veda$ wget -O minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors \$ https://huggingface.co/Veda-Sparse/Minimax-H3-T2VA-Veda-8NFE-600Step-Preview/resolve/main/minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors$ sha256sum minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors# 2a8d8845c5342756a2781e8e69563940e4bb573c9a40ebb534915ff8fd76573a
Safetensors files hold tensors only and cannot run code. Still compare the SHA256: a failed download often leaves an HTML error page saved under the model’s filename.
Use
Veda patches how MiniMax H3 computes attention. The Veda Sparse Attention node reads the predictor from models/veda and goes between the model loader and the sampler.
Veda Sparse Attention (MiniMax H3) comes from the Veda-on-ComfyUI custom node pack. Install it with ComfyUI Manager first.
MODEL
model●
Works beyond T2VA
The authors say this predictor also works for FL2VA and Ref2VA, at any step count.
Pair with Turbo
It was trained with the Turbo LoRA for 8-step generation, where the speedup is largest.
Works with
minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors is one part of the stack. A workflow that reports a missing model usually needs these too. Use the Ref2VA diffusion model instead of FL2VA for reference-to-video workflows.
| File | Role | Folder | Size | Download |
|---|---|---|---|---|
| minimax_h3_fl2va_pruned_int8_convrot.safetensors | Diffusion model (FL2VA) | models/diffusion_models/ | 20.97 GB | Get |
| qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | Text encoder (Qwen3-VL 32B) | models/text_encoders/ | 15.69 GB | Get |
| minimax_h3_video_vae_fp16.safetensors | Video VAE | models/vae/ | 5.21 GB | Get |
| minimax_h3_audio_vae_fp32.safetensors | Audio VAE | models/vae/ | 605 MB | Get |
| minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors | Turbo LoRA, 4 steps(optional) | models/loras/ | 1.96 GB | Get |
Run it
We match workflows by the model files they load: this exact file first, then other versions of it, then workflows built for the same use, then other models from the same family.
1 workflow
These workflows load another MiniMax H3 model. Select minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors in their Veda Sparse Attention (MiniMax H3) node to use it instead.
Veda Sparse Attention (MiniMax H3)
minimax_h3_r2va_veda_preview_fp8.safetensors
Replace with minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors
Compare
| File | Relation | Size | Best for |
|---|---|---|---|
| minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors | This page · T2VA preview predictor | 275 MB | All H3 modes |
| minimax_h3_r2va_veda_preview_fp8.safetensors | Sibling · trained on Ref2VA | — | Reference-to-video quality |
Troubleshoot
Why: The workflow asks for this exact filename, but ComfyUI cannot find it in models/veda. Usually the file is missing, sits in a different folder, or was saved under another name such as h3_t2va_deda_8nfe_600step_preview_fp8.safetensors.
Fix: Move the file into ComfyUI/models/veda/, rename it to match the workflow (or pick it again in the dropdown), then refresh the node list or restart ComfyUI.
Why: Veda Sparse Attention (MiniMax H3) is not built into ComfyUI. It comes from the Veda-on-ComfyUI custom node pack.
Fix: Install Veda-on-ComfyUI with ComfyUI Manager (or git clone it into ComfyUI/custom_nodes), then restart ComfyUI.
Why: The download is incomplete, or the browser saved an HTML sign-in or error page under the model’s filename.
Fix: Check that the file is exactly 275,415,648 bytes and that its SHA256 matches the value above. Download it again from Hugging Face · Veda-Sparse if not.
Rather skip the setup?
Run MiniMax H3 Veda Acceleration in your browser on a cloud GPU. Nothing to download.
FAQ
minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors, the Veda predictor on this page.models/veda. The T8 node pack uses models/veda_scorers instead. Put the file where your node looks.ComfyUI/models/veda/. Then refresh the node list or restart ComfyUI so Veda Sparse Attention (MiniMax H3) lists the file.Sources: Veda-Sparse on Hugging Face · Veda-on-ComfyUI · Veda paper (arXiv 2605.30325)
RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.