FastVideo-FastH3-Trim-Comfy Text-to-Video and Audio Model: File List

At a Glance
| Item | Value |
|---|---|
| Repository | FastVideo/FastVideo-FastH3-Trim-Comfy |
| Family guide | MiniMax-H3 guide (14 articles) |
| Publisher guide | MiniMax: models and licenses |
| Published | 2026-10-06 |
| License | minimax-h3-community-license-agreement |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
From the FastVideo team comes “FastVideo-FastH3-Trim-Comfy", a re-packaged collection of model files that makes the experimental lightweight video generation model “FastH3 Trim" available in ComfyUI. This model can generate text-to-video/audio with synchronized stereo sound in just 8 steps. Based on the “MiniMax-H3" base model, “FastVideo-FastH3-Trim-8-Step" applies weight reductions such as pruning transformer blocks, and is provided in formats like a single file (safetensors) that is easy to handle within ComfyUI workflows.
Specifications
Based on available documentation, the specifications of “FastH3 Trim" and its original model “MiniMax-H3" are as follows:
- Architecture and Structure:
- Transformer blocks reduced from 50 to 42 (the 8 blocks with the smallest impact on video and audio prediction results were removed). – Adopts rank-16 timestep conditioning. – The original model has a dense single-stream transformer core at the 33B parameter scale (H3-Omni-Transformer), utilizing the hidden states of the 50th layer of Qwen3-VL-32B as its text encoder.
- Output Specifications (Based on Original Model System Specs):
- Output Duration: 4 seconds to 15 seconds. – Aspect Ratios: Supports a wide range of aspect ratios including 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, etc. – Resolution: Shorter side defaults to 768 pixels. – Frame Rate: 24 FPS. – Audio: 32 kHz stereo. – Supported Languages (Dialogue): Stable support for 11 languages including Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.
- Recommended Settings:
- Inference Steps: 8 steps. – Sampler:
res_multistep. – Scheduler:simple. –MiniMaxH3SigmaShiftNode Settings: 10 for video generation, 3 for audio generation. - License:
minimax-h3-community-license-agreement(license notation on the model card islicense:other).
Performance and Quality
According to the model card, “FastH3 Trim" is reported to be 4.2x smaller than the base H3 model and operates faster than the previously released “FastH3 V2". However, it is explicitly stated that while block reduction is used to achieve lightweight operation and higher speed, there are trade-offs in terms of generated quality.
The base model “MiniMax-H3" has the ability to comprehensively understand text, images, video, and audio, and simultaneously predict and generate high-quality video with synchronized stereo sound. The original model’s model card publishes the following reproducible 768p generation use cases:
| Use case | Request | Result |
|---|---|---|
| T2VA | View script | t2va.mp4 |
| FL2VA | View script | fl2va.mp4 |
| Ref2VA | View script | ref2va.mp4 |
The original model supports “T2VA" for generating video and audio from text, “FL2VA" for interpolating and generating video by specifying start and end frame images, and multimodal “Ref2VA" taking images, videos, and audio as references, and is said to possess high instruction-following performance and consistency. This new “FastH3 Trim" aims to maintain these powerful generation capabilities while operating faster and more efficiently in local environments.
Strengths and Use Cases
FastH3 Trim is a model specialized for high-speed video generation in local environments. Its standout feature is the ability to execute Text-to-Video/Audio—generating video with simultaneously synchronized stereo audio from text—in an extremely short inference process of just 8 steps. It inherits the multifunctionality of the original MiniMax-H3 model, enabling video generation from 4 to 15 seconds and output in diverse aspect ratios ranging from 21:9 to 9:16 (16:9, 4:3, 1:1, 3:4, etc.).
Through experimental block reduction, it achieves a 4.2x reduction in size compared to the base model and operating speeds exceeding the preceding FastH3 V2. As a result, it is suitable not only for the latest high-end GPUs but also for trial and error on broader consumer hardware, as well as prototyping where near-real-time speeds are required. Additionally, because it stably supports dialogue generation in 11 languages (Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish), it can be leveraged for multilingual content creation.
Based on the design philosophy of the original model, it also excels at multimodal instruction-following that takes not just text, but images, video, and audio as references. This Trim version optimizes the balance between generation quality and computational cost by reducing transformer blocks from 50 layers to 42 layers while maintaining these powerful generation capabilities. It is exceptionally useful when you want to quickly generate short video clips, character movements with synchronized audio, or scenery videos.
How It Differs from Similar Models
The previously released 音声付き動画生成モデル「FastH3 V2」のFP8およびNVFP4版公開 maintained 50 transformer layers, whereas this Trim version aims for even greater speedups by reducing them to 42 layers. Furthermore, a major difference in this repository is that it is based on the model provided in 音声付き動画生成モデル「FastH3 Trim」8ステップ版が公開, but reconfigured to match ComfyUI’s standard directory structure and distributed as a single safetensors file. This allows ComfyUI users to use it simply by placing the distributed files in the designated folders without complex conversion work.
Distributed Files
Weight files published in FastVideo/FastVideo-FastH3-Trim-Comfy, listed by this site from the Hugging Face API. Sizes are the actual file sizes.
| File | Size |
|---|---|
diffusion_models/fastvideo_fasth3_trim_8step_bf16.safetensors |
37.49GB |
diffusion_models/fastvideo_fasth3_trim_8step_fp8.safetensors |
19.68GB |
diffusion_models/fastvideo_fasth3_trim_8step_int8_convrot.safetensors |
18.92GB |
diffusion_models/fastvideo_fasth3_trim_8step_nvfp4.safetensors |
11.89GB |
text_encoders/qwen3vl_32b_minimax_h3_bf16.safetensors |
51.51GB |
text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors |
27.14GB |
text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors |
15.69GB |
vae/minimax_h3_audio_vae_fp32.safetensors |
605MB |
vae/minimax_h3_video_vae_fp16.safetensors |
5.21GB |
vae/minimax_h3_video_vae_int8_convrot.safetensors |
2.81GB |
How to Get It
Model files are available from the Hugging Face repository FastVideo/FastVideo-FastH3-Trim-Comfy. Distributed in safetensors format, the following variations are provided depending on hardware generation and precision:
fastvideo_fasth3_trim_8step_nvfp4.safetensors: For NVIDIA Blackwell generation (RTX 50 series, etc.)fastvideo_fasth3_trim_8step_fp8.safetensors: For NVIDIA Ada generation and later (RTX 40 series, etc.)fastvideo_fasth3_trim_8step_int8_convrot.safetensors: For general recent NVIDIA GPUs (template default format)fastvideo_fasth3_trim_8step_bf16.safetensors: For reference quality or conversion to other formats
Usage with ComfyUI (version 0.36.0 or later) is recommended, and smooth setup can be achieved using the official FastVideo FastH3: Text to Video workflow template. Please place each acquired file in the following directories within ComfyUI’s models folder:
📂 ComfyUI/
└── 📂 models/
├── 📂 diffusion_models/
│ └── (Place fastvideo_fasth3_trim_8step_*.safetensors here)
├── 📂 text_encoders/
│ └── (Place qwen3vl_32b_minimax_h3_*.safetensors here)
└── 📂 vae/
├── minimax_h3_audio_vae_fp32.safetensors
└── (Place minimax_h3_video_vae_*.safetensors here)
During inference, it is recommended to maintain the template’s sampler settings (8 steps, res_multistep, simple) and set the MiniMaxH3SigmaShift node value to 10 for video generation and 3 for audio generation. Note that using this model requires agreement to the minimax-h3-community-license-agreement.
Can You Run It Locally?
The publisher distributes this model as safetensors.
License — minimax-h3-community-license-agreement (Commercial use prohibited): MiniMax’s MiniMax H3 Community License. Use of the weights is limited to the EU, the UK, South Korea and the US (Sections II and V.4); elsewhere, including Japan, a separate license is required.
Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Our Own Measurements
Measurements We Did Not Take
This model’s license (minimax-h3-community-license-agreement) does not permit commercial use. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).
Related Articles
- FastH3-Omni-Comfy: 80GB+ VRAM, File List
- FastH3-Omni-FP8-Comfy Audio-Synced Video Generation Model: 32GB+ VRAM
- FastH3-Omni-NVFP4-QAD-Comfy: 16GB+ VRAM, File List
- FastH3-OmniRef-Comfy: File List
What to Read Next
- Explore the same model family → MiniMax-H3 family overview (14 articles, 4 converted builds)
- Formats this model is available in → Safetensors format guide and models
- Learn about the publisher → MiniMax: models, licenses and articles

