MiniMax H 3
Offical skills to improve prompt writing: [skills on github](https://github.com/MiniMax-AI/MiniMax-H3/tree/main/skills)
The brutal truth
Audited in Stactora CMS. Architecture verified with 33.1B parameters.
Repository Sources & Packaging
This model is maintained across two distinct Hugging Face sources: the original creator's foundation repository and Comfy-Org's modular repack for local node pipelines.
MiniMaxAI/MiniMax-H3
Offical skills to improve prompt writing: [skills on github](https://github.com/MiniMax-AI/MiniMax-H3/tree/main/skills)
Comfy-Org/MiniMax-H3
Repackaged model files for ComfyUI.
Files and Versions
Direct inspected tree artifacts from Hugging Face. Verified filenames, exact byte sizes, model weight shards, and configuration definitions.
Files and versions are hosted on Hugging Face. You can view all branches, commits, model weights, and tokenizer artifacts directly.
Inspect MiniMaxAI/MiniMax-H3 /tree/main ↗Inspect /tree/main Repository ↗Easy-to-Run ComfyUI Package with Official Workflow JSON
minimax_h3_bf16.safetensorsComfyUI/models/diffusion_models/text_encoder.safetensorsComfyUI/models/text_encoders/ae.safetensorsComfyUI/models/vae/- Download weights from the official Comfy-Org Hugging Face repack.
- Place each file into its corresponding
ComfyUI/models/subfolder listed in the manifest table above. - Click "⬇ Download Workflow JSON", then drag and drop the downloaded
.jsonfile into your open ComfyUI canvas.
Decision matrix
✓ RUN / USE THIS IF…
- •You have a GPU with sufficient VRAM
- •Your 24GB GPU comfortably fits the model weights (~12GB VRAM) with dedicated headroom for inference.
- •You require 100% offline data privacy with zero third-party API logging, subscriptions, or prompt retention.
- •You build visual pipelines in ComfyUI or Automatic1111 and need repeatable, node-based diffusion workflows.
✕ WALK AWAY IF…
- •Your system lacks sufficient dedicated VRAM
- •Your GPU has less than 12GB VRAM and you cannot tolerate slow CPU/system RAM offload paging.
- •You prefer a plug-and-play web app interface and don't want to manage local model checkpoints, drivers, or runtimes.
- •You need specialized frontier reasoning that exceeds this model's parameter envelope.
True cost of ownership
About
Offical skills to improve prompt writing: [skills on github](https://github.com/MiniMax-AI/MiniMax-H3/tree/main/skills)
What it does
Download the weights from Hugging Face, pick a quantisation that matches your GPU, and run it locally with Ollama, llama.cpp, vLLM, or a compatible runtime. Image-to-video, text-to-video & audio-visual generation runs on your own hardware — no per-token bill, no third-party data handling.
Key features
Official dual-repository support: Original author MiniMaxAI and Comfy-Org modular repack
Native support for text-to-video, image-to-video, and multimodal conditioning
High temporal coherence with 39 modular safetensors weights available on Comfy-Org
Use cases
Turn a script into a polished 15–30 second clip.
Visualise shot lists before a full production run.
Fill timeline gaps with convincing atmospheric footage.
Create stylised clips from a lyric sheet or mood board.
Quantization & Hardware Fit
VRAM Hardware Compatibility & Memory Allocation
| Quant | VRAM | Quality / Fit on 24GB | Install / Run Command (COMFYUI) | |
|---|---|---|---|---|
| Q2_K (2-bit Extreme)● ACTIVE | 12 GB | 55% quality retention · ✓ Fits (12GB spare) | comfy model download --url https://huggingface.co/Comfy-Org/MiniMax-H3 --relative-path ComfyUI/models/diffusion_models/ | |
| Q3_K_M (3-bit Compact) | 17 GB | 68% quality retention · ✓ Fits (7GB spare) | comfy model download --url https://huggingface.co/Comfy-Org/MiniMax-H3 --relative-path ComfyUI/models/diffusion_models/ | |
| Q4_K_M (4-bit Sweet Spot) | 22 GB | 82% quality retention · ✓ Fits (2GB spare) | comfy model download --url https://huggingface.co/Comfy-Org/MiniMax-H3 --relative-path ComfyUI/models/diffusion_models/ | |
| Q5_K_M (5-bit High Precision) | 26 GB | 91% quality retention · ✕ Needs 26GB | comfy model download --url https://huggingface.co/Comfy-Org/MiniMax-H3 --relative-path ComfyUI/models/diffusion_models/ | |
| Q6_K (6-bit Studio) | 31 GB | 96% quality retention · ✕ Needs 31GB | comfy model download --url https://huggingface.co/Comfy-Org/MiniMax-H3 --relative-path ComfyUI/models/diffusion_models/ | |
| Q8_0 (8-bit Reference) | 40 GB | 99% quality retention · ✕ Needs 40GB | comfy model download --url https://huggingface.co/Comfy-Org/MiniMax-H3 --relative-path ComfyUI/models/diffusion_models/ | |
| F16 (16-bit Unquantized) | 74 GB | 100% quality retention · ✕ Needs 74GB | comfy model download --url https://huggingface.co/Comfy-Org/MiniMax-H3 --relative-path ComfyUI/models/diffusion_models/ |
Buyer doubts, cleared
Can I actually run MiniMax H 3 on my 24GB GPU?
What's the fine print on the license?
Does the Image/Text to Video / ComfyUI actually hold up in practice?
User reviews
The Q2_K (2-bit Extreme) build is the sweet spot. Anything larger needs serious hardware for marginal quality gains.
Used MiniMax H 3 for a client project. Documentation was solid, though the fine-tuning path is less polished than base inference.
We evaluated MiniMax H 3 against two hosted alternatives. It won on cost-per-output but lost on latency. For batch work it is a clear yes.
The output reads naturally once you tune the prompt. It has cut my first-draft time roughly in half.
