Check a model release and find its official tools
33B open weights, picture and audio generated together. Every link, the hardware reality, and the parts that aren't open.
By Alan Brown
MiniMax H3 is a 33B open-weights video model that generates 24fps picture and 32kHz stereo audio together in one pass, rather than dubbing audio on afterwards. Launched 31 July 2026; base weights opened on 3 August, the same day ComfyUI shipped native support.
- 1️⃣ Official weights — huggingface.co/MiniMaxAI/MiniMax-H3. The real thing: 33B omni-modal, with FL2VA (text / first-last-frame) and Ref2VA (reference) checkpoints in Diffusers format.
- 2️⃣ ComfyUI repack, start here for local — huggingface.co/Comfy-Org/MiniMax-H3. Pruned INT8 files for ComfyUI's native nodes: about 42.5GB against 123.6GB at full precision. Needs ComfyUI 0.30.0+.
- 3️⃣ Setup guide — blog.comfy.org/p/minimax-h3-day-0-support-in-comfyui. The day-0 support post: six official workflow templates plus the memory-offloading explanation.
- 4️⃣ Low-VRAM builds — huggingface.co/Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot. Community INT4 / INT8 / NVFP4 pack with a pick-by-your-GPU guide. This is the one for smaller cards.
- 5️⃣ Code and deployment scripts — github.com/MiniMax-AI/MiniMax-H3
- 6️⃣ Try it without installing anything — platform.minimax.io for the official API and playground, or fal.ai/minimax-h3 for a one-click test.
- 7️⃣ The announcement — www.minimax.io/news/minimax-h3-open-source
- 🧬 Architecture: a 33B dense Omni Transformer with a Qwen3-VL-32B encoder. Picture and sound are generated jointly, not stitched together.
- 🎛 Inputs: up to 9 reference images, 3 video clips and 3 audio tracks in a single generation. It reads identity, performance, camera movement, composition and editing rhythm off the references and carries them through.
- ✂️ Editing: swap a product, rewrite signage, relight day to night, add or remove objects — while the rest of the shot holds still. It renders legible type, subtitles and brand marks, and will clone a voice from a reference recording onto your character.
- 🖥 Hardware: Comfy pruned roughly 40% of the modulation weights into a lookup table and quantised to int8 — 123.6GB down to 42.5GB. Clips still take several minutes on a 3090.
- 💵 Hosted pricing: $0.13 per second at 2K (about $7.80 a minute), with a 768p tier at $0.09/s in closed beta.
- 🏷 Naming: H3 is the model. Hailuo 3.0 / Hailuo 03 is the app-side alias and the API endpoint is labelled Hailuo-03. Nothing to do with Kling O3.
Context & limitations
Downloading the weights does not give you Hailuo. H3-Context-IR (the prompt preprocessing) and H3-Regenerate-2K (the upscale) are API-only, so a local install reproduces the model and not the pipeline the hosted product actually runs.
The audio is the weak part, which is awkward for a model whose headline is joint audio. Native audio is unreliable for scripted dialogue — testers got repeated syllables and unrelated audio on several clips. Treat every generation as a draft and listen to it. Picture realism and shot arrangement still have gaps too.
The leaderboard line is narrower than it sounds: #1 on Artificial Analysis for audio-enabled video EDITING at an Elo of 1130, ahead of Gemini Omni Flash, HappyHorse-1.0 and Wan 2.7. It's #2 in text-to-video and #3 in image-to-video. Say "editing", or the number is wrong.
And read the licence before self-hosting. It's the MiniMax H3 Community License, and the model card carries an application form specifically for the USA, EU, UK and South Korea — those four are a separate track rather than a straight download. Commercial use is royalty-free otherwise, but above US$20M yearly revenue needs written authorisation, products have to show "MiniMax H3" in the UI, and outputs and weights can't be used to train other models. Your outputs are yours.
Do this yourself
Exact model cards and workload settings are more reliable than a fixed list of download sizes.
Want to build this with a teacher?
Pathwise Academy runs small, hands-on AI classes in Singapore for ages 10–17 and adults.