MiniMax H3 Video Online — No GPU, No 42.5GB Download

MiniMax H3 is a 33B-parameter omni-modal video model from MiniMax AI, announced July 31, 2026. Omni-modal here means one model produces picture and sound in the same pass rather than generating video and dubbing audio onto it afterwards. It generates clips of 4 to 15 seconds at up to 2K and 24fps, with native stereo audio, and supports text-to-video, first-frame, last-frame, and combined first-and-last-frame control, plus instruction-based editing of existing footage. On the Hailuo question: H3 is the successor to the Hailuo series. Hailuo was MiniMax's video product branding through the 02 generation, so hailuo minimax h3 searches are usually people checking whether the two names point at the same lineage — they do, with H3 as the newer generation. If you are looking for minimax m3 instead, note that MiniMax's M-series naming covers its language models, a separate product line from the video work.

The weights are on Hugging Face, published August 3, 2026, with the project repository on GitHub — roughly 42.5GB for the checkpoint before encoders and support files. For hardware: it is a 33B model, so realistic minimax h3 vram requirements start around a 32GB-class card, with a 5090 the commonly cited practical floor for full-resolution work. Community int8 and GGUF quantisations bring that down at some cost in fidelity, and minimax h3 mac and AMD support depend on backend maturity rather than on the weights themselves. Two caveats before you start the download. First, the MiniMax Community License restricts local deployment rights in several regions, covered in the next answer. Second, some modules of the full 2K pipeline have not been open-sourced, so a local install reproduces a subset of what the official demos show — the workflow will run, but the top-tier 2K output is not currently reproducible outside the official pipeline.

Not if you are in the United States, the European Union, the United Kingdom, or South Korea. The MiniMax Community License that ships with the H3 open weights excludes those regions from the local deployment grant, which is a licensing restriction rather than a technical one — the download will work, the permission to run it on your own hardware is what is missing. Hosted access is unaffected. The official API and platforms built on it remain available globally, because MiniMax operates the weights on its own infrastructure and licenses the service rather than the deployment. That is the practical route for teams in restricted regions, and it is what the panel at the top of this page uses. As always with licence terms: this reflects the LICENSE file as published and checked on August 12, 2026, it is a summary rather than legal advice, and terms get revised — read the official repository, and talk to your own counsel if the answer carries real commercial weight for you.

MiniMax's published minimax h3 pricing for the API is $0.13 per second of generated video at 2K, which works out to $7.80 per minute. A cheaper 768p tier at $0.09 per second has been listed as coming soon rather than live. On references: the first 5 reference images are free, additional ones are $0.04 each, and reference audio carries no charge. Because clips are 4 to 15 seconds, a single 2K generation lands somewhere between roughly $0.52 and $1.95 before reference-image costs. Worth noting what you get for that: the API path runs the complete 2K pipeline, including the modules that were not open-sourced, so it is not simply a paid convenience layer over what you could self-host. Prices are as published at the time of writing, August 12, 2026, and API pricing moves — check MiniMax's own pricing page before budgeting against these numbers. The panel on this page bills in credits instead, with free trial credits on new accounts. For a side-by-side against the other metered option people weigh this against, xAI publishes $0.04 per image and $0.080 per second of video on the Grok Imagine API — the API pricing comparison page works through what each unit actually costs on a real workload.

ComfyUI native support for H3 was merged on August 3, 2026, the same day the weights were published, and official workflow templates shipped with it — so a minimax h3 comfyui workflow does not have to be assembled from scratch. The practical path is to update ComfyUI to a build that includes the merge, pull the weights from Hugging Face, place them where the template expects, and load the official workflow. Expect a large download and a first-run that spends a while compiling and loading. Two things to set expectations on. The regional licence restriction applies here in full — comfyui minimax h3 setups in the US, EU, UK, or South Korea run into the same local deployment limit as any other install, and ComfyUI supporting the model does not change what the licence permits. And the unreleased 2K pipeline modules mean the ComfyUI output ceiling sits below the official demos. If either of those is a blocker, the online panel above skips both.

As of August 12, 2026, MiniMax has not announced an official Turbo checkpoint, a LoRA training pipeline, or a separately branded Director model for H3. That matters because minimax h3 turbo lora and minimax h3 director are actively searched terms with very little published behind them yet. Some of that is pattern-matching from earlier releases: MiniMax's Hailuo generation did ship a Director-branded variant aimed at camera-movement instructions, so people reasonably expect the equivalent here. For H3, camera control lives in the prompt instead — shot type and movement described in plain language, which the model handles well. Community fine-tunes and speed-optimised builds typically appear within weeks of any open-weight release, and H3 will likely be no exception, but anything you find today is third-party, unvalidated, and inherits the same licence restrictions as the base weights, including the regional deployment clause.

Short version: as of the Artificial Analysis boards in August 2026, H3 leads on video editing, sits second on text-to-video, and third on image-to-video, which puts it in the same top tier as the current Seedance and Kling releases rather than clearly ahead of them across the board. Its clearest edge is instruction-based editing and native stereo audio in a single pass; where it gives ground is clip length, since H3 caps at 15 seconds. Which one wins depends entirely on the shot you are making, and leaderboard positions shift as models ship. We have dedicated comparison pages that go through this properly rather than in a paragraph — see the Seedance and Kling comparisons linked below. The genuinely useful move, if you have credits on this site, is running the same prompt through two models and judging the output rather than the ranking.

The minimax h3 reddit threads after the August 3 weights release split roughly three ways. The instruction-based editing drew the most positive attention, which tracks with the Artificial Analysis ranking. VRAM was the dominant complaint — a 33B model plus a 42.5GB download rules out most consumer setups, and quantisation threads appeared almost immediately. The third theme, and the one that took a few days to surface, was the licence: the regional deployment restriction was not widely noticed until people started reading the LICENSE file, and it changed a lot of local-install plans in the US and EU. Standard caveat on all of this — early launch reactions are noisy, run heavily on first impressions, and skew toward whoever had hardware to test on day one. Useful as direction, not as verdict.