Now, the next Open sourced Video generation model is here. MiniMax H3 (Developed by MiniMaxAI) is positioned as a next-generation
general-purpose multimodal video model designed to move beyond
specialized generation tasks toward broader multimodal intelligence.
Rather than limiting itself to text prompts, can understand a
combination of Text,Images,Video and Audio.
The idea is
straightforward is to give the model more context, and it can better
understand what the creator actually wants. The model can generate
videos with native stereo audio, supporting resolutions up to 2K and
durations of up to 15 seconds.
It is also designed for commercial scenarios such as film and television, advertising, branding, e-commerce, and gaming.
 |
| minimax H3 full model architecture |
One of the most interesting aspects of H3 is that it is not simply a single generation model. The complete system is divided into three interconnected modules. MiniMax H3 is released under the MiniMax H3 community licensing terms. Commercial users should review the license carefully, particularly if their products or services generate more than US $20 million in revenue, as separate written authorization from MiniMax may be required under the stated terms.
 |
| minimax H3 base architecture |
When a prompt contains multiple types of information, understanding everything correctly becomes difficult. H3-Context-IR acts as the system's context-processing layer. It takes complex multimodal instructions and converts them into a structured representation called Context Intermediate Representation (Context-IR).
This representation
gives the generation model a clearer understanding of the user's intent.
The Context-IR stage is particularly important because better
interpretation of the original inputs directly affects the quality of
the final video.
Once the context has been processed, H3-Base
generates the video and audio. The initial generation is produced at
768p resolution. This stage focuses on turning the multimodal context
into a coherent audiovisual sequence.
The system then takes the
generated 768p result along with the original context and sends them
back through H3 for another generation pass. This produces the final 2K
output. The benefit is that the model has access to both the initial
generation and the original multimodal instructions, allowing it to
improve details and visual fidelity instead of simply upscaling the
existing video.
Installation
1. Make sure you have the latest ComfyUI installed on your system. Older user, just need to update ComfyUI from the Manager.
2. Download Minimax H3 model from Hugging face repository. There are multiple variants listed below (for T2V, I2V, Ref2V) released by the community, download any of them as per your system resources:
(A) Download Minimax H3 model (repacked by official ComfyUI)
| Model Name | Type | Description | Size | VRAM Needed |
|---|---|---|---|---|
| minimax_h3_fl2va_bf16.safetensors | BF16 | Full-precision FL2VA model designed for high-quality video generation with maximum numerical precision. | 66.3 GB | 80+ GB |
| minimax_h3_fl2va_int8_convrot.safetensors | INT8 | INT8 FL2VA model with ConvRot quantization, reducing memory usage while retaining strong generation quality. | 34 GB | 40+ GB |
| minimax_h3_fl2va_pruned_int8_convrot.safetensors | Pruned INT8 | Pruned INT8 FL2VA model with ConvRot quantization, optimized for lower VRAM usage compared with the full INT8 model. | 21 GB | 24–32 GB |
| minimax_h3_ref2va_bf16.safetensors | BF16 | Full-precision REF2VA model designed for high-quality video generation with maximum numerical precision. | 66.3 GB | 80+ GB |
| minimax_h3_ref2va_int8_convrot.safetensors | INT8 | INT8 REF2VA model using ConvRot quantization to reduce memory requirements while maintaining strong generation quality. | 34 GB | 40+ GB |
| minimax_h3_ref2va_pruned_int8_convrot.safetensors | Pruned INT8 | Pruned INT8 REF2VA model with ConvRot quantization, optimized for systems with more limited GPU memory. | 21 GB | 24–32 GB |
| minimax_h3_fl2va_pruned_fp8_scaled.safetensors | Pruned FP8 | Pruned FP8 FL2VA model with scaled quantization, offering a balance between generation quality and reduced VRAM usage. | 21 GB | 24–32 GB |
| minimax_h3_ref2va_pruned_fp8_scaled.safetensors | Pruned FP8 | Pruned FP8 REF2VA model with scaled quantization, designed for efficient inference with substantially lower memory requirements. | 21 GB | 24–32 GB |
*Note:
REF2VA - Reference to Video+Audio
FL2VA - First Frame last Frame to Video+Audio
Save any of them into your ComfyUI/models/diffusion_models folder.
(B) Download Minimax H3 model INT8/INT4/NVFP4 mixed(Community quantized compiled)
| Model Name | Type | Description | VRAM | Size |
|---|---|---|---|---|
| MiniMax_H3_FL2VA_pruned_mixed_int4_int8_convrot.safetensors | Mixed INT4/INT8 | Mixed-precision pruned model that combines INT4 and INT8 quantization for a balance between memory efficiency and generation quality. | 15+ GB | ~15.5 GB |
| MiniMax_H3_Ref2VA_pruned_mixed_int4_int8_convrot.safetensors | Mixed INT4/INT8 | Mixed-precision pruned model combining INT4 and INT8 quantization, designed to reduce VRAM usage while maintaining good output quality. | 15+ GB | ~15.5 GB |
| MiniMax_H3_FL2VA_pruned_int8_convrot.safetensors | INT8 | Pruned INT8 model optimized for high-quality generation while significantly reducing memory requirements compared with the full-precision version. | 24+ GB | ~21 GB |
| MiniMax_H3_Ref2VA_pruned_int8_convrot.safetensors | INT8 | Pruned INT8 model designed for high-quality generation with lower memory usage than the full-precision REF2VA model. | 24+ GB | ~21 GB |
| MiniMax_H3_FL2VA_pruned_int4_convrot.safetensors | INT4 | Highly compressed pruned model designed to run on standard consumer GPUs with significantly lower VRAM requirements. | 12–16 GB | ~11.3 GB |
| MiniMax_H3_Ref2VA_pruned_int4_convrot.safetensors | INT4 | Highly compressed REF2VA model that reduces memory consumption while remaining suitable for consumer-grade hardware. | 12–16 GB | ~11.3 GB |
| MiniMax_H3_Ref2VA_pruned_nvfp4.safetensors | NVFP4 | Experimental Nvidia 4-bit floating-point model optimized for Blackwell-generation GPUs. | 16+ GB | 12.5 GB |
| MiniMax_H3_Ref2VA_nvfp4_mixed.safetensors | NVFP4 Mixed | Experimental mixed NVFP4 model using Nvidia 4-bit floating-point quantization. Requires a Blackwell architecture GPU. | 24+ GB | 24.4 GB |
| MiniMax_H3_FL2VA_pruned_nvfp4.safetensors | NVFP4 | Experimental pruned Nvidia 4-bit floating-point model designed for efficient inference on Blackwell architecture GPUs. | 16+ GB | 12.5 GB |
Save any of them into your ComfyUI/models/diffusion_models folder.
(C) Download Minimax H3 GGUF (By Rebel) or Minimax H3 GGUF (by vantageAI)variant. There are Q2(fast inference with quality degrade) to Q8(good quality with slow inference) model variants available.
| Model Name | Type | Description | Size | VRAM Needed |
|---|---|---|---|---|
| MiniMax-H3-FL2VA-Q2_K-(Mixed_Precision).gguf | Q2_K Mixed Precision | Highly compressed mixed-precision GGUF model using Q2_K quantization, designed to significantly reduce memory usage while providing efficient MiniMax H3 FL2VA inference. | 18.7 GB | 16+ GB |
| MiniMax-H3-FL2VA-Q3_K_M.gguf | Q3_K_M | 3-bit medium-quantized GGUF model that provides a balance between model size, VRAM efficiency, and generation quality. | 15.6 GB | 16+ GB |
| MiniMax-H3-FL2VA-Q4_K_M.gguf | Q4_K_M | 4-bit medium-quantized GGUF model offering a stronger quality-to-memory balance while requiring more storage and VRAM than lower-bit variants. | 19.9 GB | 20+ GB |
Save any of them into your ComfyUI/models/unet folder. Make sure you have ComfyUI-GGUF custom node (by city-96) already installed.
3. Download any of the text encoders.
(a) qwen3vl_32b_minimax_h3_bf16.safetensors (for 32 GB Vram)
(b) qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (For Blackwell 16 gb Vram)
(c) qwen3vl_32b_minimax_h3_int8_convrot.safetensors (Recommended for 24GB Vram)
Save this into ComfyUI/models/text_encoders folder.
4. Download both of the VAEs as they are required for audio-video encoding /decoding.
(a) minimax_h3_audio_vae_fp32.safetensors
(b) minimax_h3_video_vae_fp16.safetensors
Save this into ComfyUI/models/vae folder.
5. (Optional) Download Minimax H3 Turbo LoRA or use this Minimax H3 Turbo lora (same converted for comfyUI). Save this into ComfyUI/models/loras folder. Its renders joint video and synchronized stereo audio in 6-8 sampling steps. If using lora models, use the Load Lora node between Load diffusion model loader/unet loader node and the Guider/Sampler nodes.
6. Restart and Refresh ComfyUI to take effect.
Workflow
1. Download the Minimax H3 workflows. This supports with three types:
(a) Minimax_H3_img2video.json (image to Video workflow)
(b) Minimax_H3_ref2video.json (Reference to Video workflow)
(c) Minimax_H3_txt2video.json (Text to Video workflow)
If using GGUF variants, just replace the load diffusion model node with unet loader node when working with the workflows.
2. Drag and drop into ComfyUI. The model supports videos between 4 and 15 seconds at 24 FPS. It also supports a wide range of aspect ratios, including: 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16.
It also provides stable support for dialogue in 11
languages such as English, French, Arabic, Chinese, Korean, Portuguese,
Russian, Japanese, Spanish, German, Italian. You can follow official Minimax H3 prompting guide to get the best results.
Image to Video
Text to Video
Reference to Video
){kind=link}