MiniMax H3 (also called Hailuo 3) is an open-weight video model: MiniMax published the weights on Hugging Face under the MiniMax H3 Community License. That means you can run it yourself — but it's a large model, and the hardware decides how much of it you actually get. This guide sums up what the model card and community guides say, so you can decide between running it locally and using it online.
What you're downloading
According to the model card, H3 is a 33-billion-parameter, single-stream Transformer (MiniMax calls it the H3-Omni-Transformer) that uses Qwen3-VL-32B as its encoder. The released checkpoints are CFG-distilled and run in BF16. It generates 4 to 15 seconds at 24 fps with 32 kHz stereo audio; the shorter side is 768 pixels by default, and 2K comes from a separate in-context regeneration step (H3-Regenerate-2K).
For a local setup you typically need four pieces, as community ComfyUI guides list them: the H3 denoiser, the Qwen3-VL text encoder, the video VAE and the audio VAE.
Before you build anything on the weights, read the MiniMax H3 Community License on the model page — it sets its own terms of use.
Ways to run it
The model card names four runtimes: SGLang, vLLM, diffusers and ComfyUI.
- SGLang / vLLM — for servers. MiniMax's own SGLang example runs on 4 GPUs (
--num-gpus 4,--ulysses-degree 4) with--performance-mode speed. - diffusers — for Python scripts and custom pipelines.
- ComfyUI — the usual choice on a desktop GPU, with node workflows for text, image and reference modes.
How much VRAM do you need?
The full BF16 model is meant for multi-GPU servers. On a single consumer card, people run quantized (GGUF, FP8 or INT8) builds, which trade quality for memory. Community guides give roughly this picture:
| VRAM | What to expect |
|---|---|
| 8–12 GB | Smallest quantized builds; experiments only, short or low-resolution clips |
| 16 GB | The realistic entry point for quantized tests |
| 24 GB+ | Better quantization levels if the whole workflow loads |
| 32 GB+ | Higher-quality experiments, still not the full model |
System RAM and a fast NVMe drive matter too, because large parts of the model are swapped in and out. Treat these numbers as community reports, not guarantees — they change with every new workflow and quantization.
Local vs online: which is cheaper?
Run it locally if you already own a 24 GB+ GPU, want to fine-tune or use LoRAs, or need to keep everything offline.
Use it online if you want the full model (not a quantized build), 2K output, reference images, clips and audio together, and no setup. On MiniMax H3 AI you pay only for the seconds you render — H3 Max starts at 6 credits a second at 480P, so a 5-second clip is 30 credits, and the smallest pack is 300 credits for $9.99 — and failed renders give the credits back. Try it in the browser.
Prompts work the same either way
Whether you run H3 in ComfyUI or online, the model reads the same prompts. Our MiniMax H3 prompt guide covers the structure MiniMax recommends — the timeline, camera moves, dialogue and sound — with examples you can copy.
Sources
- MiniMax H3 model card on Hugging Face — architecture, durations, frame rate, audio, runtimes, license.
- MiniMax H3 announcement — capabilities and release.
- Community ComfyUI and GGUF guides for the VRAM ranges above.
Read more
MiniMax H3 vs H3 Max: Which One Should You Use?
Resolution, length, speed and price of MiniMax H3 and its fast variant H3 Max, and when to pick each.
MiniMax H3 Prompt Guide: The Official Structure, Camera Moves and Examples
How to prompt MiniMax H3 (Hailuo 3): a simple five-part structure, the three-field format from MiniMax's own prompting guide, camera-move vocabulary, multi-shot timing, dialogue and sound — with examples you can copy.