MiniMax open-sourced its H3 video model on August 3, 2026, and it is the most consequential video-weights drop of the year: a general-purpose multimodal generator that reads text, images, video and audio as one context and produces 4–15 second clips at up to 2K, 24fps, with native 32kHz stereo audio — sound generated with the picture, not dubbed after it. The weights are live on Hugging Face and ModelScope, ComfyUI support shipped day-zero, and the community had it running on a single consumer GPU within hours. Then there is the license paragraph nobody should skip: the agreement's applicable territory excludes the European Union, the United Kingdom, South Korea and the United States. The most capable open video model to date is open for most of the world — and formally off-limits in most of the Western market reading about it.
What actually shipped
The headline number is smaller than the hype suggested, and that is the good news. H3 is a 33B-parameter dense single-stream Transformer — not the 100B+ monster early rumors described — of which roughly 13B sits in AdaLN branches that can be cached at inference. Full BF16 weights run 123.6GB; a pruned int8 build lands at 42.5GB.
The system is three modules, and only the middle one is fully yours:
- H3-Context-IR — the multimodal context preprocessor. Not open-sourced; local workflows substitute standard encoders.
- H3-Base — the generator itself, natively 768p (short side). This is what the open weights are.
- H3-Regenerate-2K — the in-context upscaler behind the 2K marketing number. The official model card lists it as not yet open-sourced; 2K output today is effectively an API-side capability.
Two task checkpoints cover the workflow split: FL2VA (text-to-video, first/last-frame control) and Ref2VA (reference-driven generation from multimodal input). Aspect ratios span 21:9 to 9:16; prompts hold across 11 languages. One detail worth savoring: the text encoder is Qwen3-VL-32B under Apache 2.0 — Alibaba's model doing the reading inside MiniMax's video stack. The Chinese open ecosystem is not just shipping in parallel; it is composing.
Can you actually run it?
Yes — with caveats that are honest rather than fatal. A community tester ran the ComfyUI build on an RTX 5070 Ti (16GB) with 48GB of system RAM: about 3 minutes for a 5-second clip, 8 minutes for 10 seconds, VRAM pinned at 98%, via the official ComfyUI workflow described as having "basically no barrier to entry." ComfyUI claims the pruned int8 checkpoint runs on 12GB cards with dynamic offloading and ~64GB system RAM — their assertion, not an independently tested one. Serious throughput still means datacenter cards, but "runs overnight on a gaming PC" is now literally true for short clips, and that is a first for a model of this quality.
The early quality consensus from community testing: base output lands around the level of the leading commercial generators' previous tier — with the crucial difference that a local model enforces no content policy beyond the license text, which is precisely why the export-cost math for short-form video studios is being redone in public threads.
The license has teeth
Read before building:
- Territory: the Community License names the EU, UK, South Korea and the United States as excluded. MiniMax's own QA attributes this to video generation's "complex and rapidly evolving regulatory environment" — plausibly AI-Act-shaped caution. Whatever the reason: if your company operates in those jurisdictions, local deployment is not licensed, full stop. API access through providers remains a separate, unaffected channel.
- Attribution: commercial use is free, but your product's UI must display "MiniMax H3."
- Scale gate: above $20M in annual revenue, you need separate written authorization — the same MaaS-style gate Kimi K3's license normalized in July. "Open weights, tiered rights" is now simply how Chinese labs ship.
Where this fits
August 3 was the day Chinese open-source fired on both barrels: Qwen3.8-Max went official with a weights commitment in text, and MiniMax delivered actual downloadable weights in video — the medium where open alternatives have lagged commercial quality the longest. The community line that MiniMax should "stick to what it's great at instead of fighting the coding war" lands as vindicated: H3 is the lab playing its strongest suit.
For developers, the practical decision tree is unchanged in shape and updated in numbers:
- Self-hosting H3 buys you unlimited generations at hardware-and-electricity cost, full content control, and license risk you must actually read (territory, attribution, revenue gate) — plus 768p native output until the 2K module opens.
- Hosted video APIs buy you 1080p pipelines, no VRAM management, per-second billing with failed generations refunded, and zero license exposure. On Turiloop, the HappyHorse lineup serves audio-native 1080p from $0.14/second behind the same OpenAI-compatible key as the text models — the full API walkthrough takes about ten minutes to first clip.
The honest rule: prototype and burst on an API; consider self-hosting when volume is steady, your jurisdiction is licensed, and 768p fits the product. H3 just made the second option real for video — for most of the world.
FAQ
What is MiniMax H3? A general-purpose multimodal video generation model open-sourced by MiniMax on August 3, 2026: 33B dense parameters, generating 4–15 second clips up to 2K/24fps with native 32kHz stereo audio from text, image, video and audio context. Weights are on Hugging Face and ModelScope with day-zero ComfyUI support.
Can I use MiniMax H3 in the United States or Europe? Not under the open-weights license — the Community License's applicable territory explicitly excludes the US, EU, UK and South Korea. Using the model via hosted APIs is a separate channel governed by provider terms, not the weights license.
What hardware does MiniMax H3 need? The pruned int8 checkpoint is 42.5GB; community testing ran it on an RTX 5070 Ti (16GB VRAM, 48GB RAM) at roughly 3 minutes per 5-second clip via ComfyUI. ComfyUI claims 12GB cards work with offloading and ~64GB system RAM. Full BF16 weights (123.6GB) and real throughput remain datacenter territory.
Is MiniMax H3 free for commercial use? Yes within the licensed territory, with two conditions: your UI must display "MiniMax H3," and annual revenue above $20M requires separate written authorization from MiniMax.
Does MiniMax H3 output 2K? The 2K number comes from the H3-Regenerate-2K upscaling module, which the official model card lists as not yet open-sourced. The open H3-Base generates at 768p (short side) natively; 2K is currently an API-side capability.
H3 vs HappyHorse — which for production video? H3 is the self-hosting play: unlimited volume at hardware cost, 768p, license permitting. HappyHorse is the hosted play: audio-native 1080p, per-second billing from $0.14/s, failed tasks auto-refunded, no infrastructure. Prototype hosted; self-host only with steady volume and a licensed jurisdiction.