MiniMax H3: how it works, and what you can actually run

ComfyUI's tutorial page for MiniMax H3 says two things about resolution. Near the top: output is up to 2K, 24fps, about 15 seconds. Further down, describing what the model does: H3's native canvas is a 768px short edge, capped at 768x1344 pixels. Both sentences sit on the same page. Both are accurate, because they describe different software.

MiniMax launched H3 on 2026-07-31 and released the weights three days later, on 2026-08-03. In between, it took first place in video editing on Artificial Analysis and the coverage settled into a shape: an open-weights omni-modal model, 2K with native stereo audio, one third the price of the closed frontier, runs on an RTX 3060.

What downloads is one stage of a three-stage system, and the licence attached to it excludes the European Union, the United Kingdom, the Republic of Korea and the United States. Everything below is as of 2026-08-04, which for a four-day-old release matters more than usual.

MiniMax shipped one third of the system it described

The phrase "the complete H3 system" is MiniMax's own. Its model card names three modules.

H3-Context-IR reads the multimodal instructions and converts them into an intermediate representation. MiniMax calls it "critical to the quality of the final output" and then says it "is not included in this open-source release." H3-Base generates video and audio at 768p from that representation. This is the part that shipped. H3-Regenerate-2K takes the 768p result and the original context and regenerates at 2K, and MiniMax says it "is not yet open-sourced. We will release it once it is ready."

A third thing was held back and got almost no attention. The model was trained with native sparse attention, and the model card is blunt about what that means for the download: "The initial open-source release provides inference with full attention only." The artifact is lower-resolution than the demos and slower than the production system.

None of this is hidden. MiniMax states each omission plainly in its own documentation, which is more than several vendors manage. The gap is between the documentation and the framing: the launch post is titled "An Open Model Breaking the Boundaries Between Tasks and Modalities", and its own closing line promises a technical report that, four days later, has not appeared.

Worth listing what else has not appeared, because for a release framed as open the absences are the story. There is no technical report. There is no arXiv paper. There is no GitHub repository, which breaks MiniMax's own pattern for every M-series model it has published. There is no training-data disclosure of any kind, and no benchmark table.

The architecture is one sequence, and that is the interesting part

Most video models generate pictures and then solve audio separately. H3 does not.

The diffusers integration, written against the real checkpoint, describes it precisely: one transformer denoises a single packed sequence holding the text conditioning, the conditioning image, the video and audio rows, and the target rows, with full self-attention over all of it. "There is no separate vocoder and no audio post-hoc pass: video and audio come out of the same denoising loop."

Getting four modalities into one sequence takes three encoders. Text goes through H3-Encoder, which is the full pretrained Qwen3-VL-32B read at its 50th layer rather than its last, with the language-model head unused. Visual input goes through both H3-Encoder and H3-VisualVAE, a temporally causal autoencoder at 16x spatial and 4x temporal compression with 24 latent channels, then patchified to an effective 32x spatial downsample. Audio goes through H3-AudioVAE alone, which compresses 32 kHz down to 40 Hz of latent tokens per channel and runs the same encoder independently on left and right, which is how stereo works at all. Three-dimensional rotary position embeddings carry the relationships across time and the two spatial axes.

The transformer itself is deliberately plain. It is 33B parameters, dense, single-stream, with roughly 13B of that sitting in AdaLN branches whose outputs can be precomputed and cached. Neither the attention layers nor the feed-forward layers contain anything modality-specific.

Then the part that matters for everything else in this article. MiniMax did not build an upscaler:

For H3's 2K-resolution output, instead of using a conventional dedicated super-resolution module, we use the H3 base model to regenerate its own low-resolution result through an in-context manner.

The reasoning is sound. A super-resolution network guesses at detail it never saw. Regenerating in context lets the model re-read the original instructions, so small text and fine detail can be recovered rather than invented.

It also has a consequence. A super-resolution network is a file, and files get released. A pass that re-reads the original context through the base model is a pipeline, and a pipeline is a service. MiniMax's stated reason for withholding it is "the complexity of the system," which is exactly what a pipeline is.

That reading has a limit, and it should be stated rather than glossed. MiniMax never says H3-Regenerate-2K runs the identical weights you downloaded with no additional trained parameters. "Complexity of the system" points at extra machinery beyond a weight file. Whether the 2K stage is your checkpoint plus orchestration, or your checkpoint plus something else, MiniMax has not said, and nobody outside the company can currently tell.

What is settled is that no local path to 2K exists. The "Full 2K Workflow" script shipped in the repository posts to $MINIMAX_API_BASE/v2/video_regeneration, while the 768p script posts to a local server. The model card describes the workflow as "combining the Open Platform API with a locally deployed H3-Base". Someone asked about local 2K upscaling in the Comfy-Org discussions on 2026-08-03 and received no reply.

# MiniMax's own reproducible 768p request, local
curl http://localhost:30010/v1/videos \
  -d '{ ... "target": {"short_edge": 768, ...} }'

# MiniMax's own 2K step, remote
curl "$MINIMAX_API_BASE/v2/video_regeneration" \
  -H "Authorization: Bearer $TOKEN" \
  -d '{ ... "resolution": "2K" }'

One correction to the framing that has settled into the coverage, including an earlier version of my own reading. The 768px short edge is a default and a trained canvas, not a hard cap. Height and width only have to be multiples of 32, and people have pushed well past it: a ComfyUI issue reports a run at 1920x1088, and one Hacker News user set the megapixel control to 2.0 and got the same. Quality is the constraint, not the code. A user on a larger machine tried a 10-second 1080p clip and reported "the results were quite poor. Unusable for anything, in fact."

The leaderboard result is real and it does not describe your download

Artificial Analysis runs blind pairwise human preference pools for video. Read live on 2026-08-04, its boards place H3 as follows.

Top three in all six pools is true, and it is a genuinely strong showing for an open-weights model. The headline underneath it needs a condition that the launch coverage drops: first place in video editing holds in the with-audio pool only. Strip the audio and H3 falls to third, inside what Artificial Analysis itself marks as a statistical tie for second. H3's editing lead is substantially an audio lead.

Two more things the boards say that the coverage does not. Gemini Omni Flash beats H3 in five of six pools, and on Artificial Analysis's own normalized pricing column it is cheaper, $6.00 per minute against $7.80. And the model being voted on is the hosted 2K API with H3-Context-IR in front of it. Not the 768p weights.

No VBench, VBench-2.0, VideoPhy or Movie Gen Bench result for H3 exists from anyone. A VBench score of 83.41 circulating in search results belongs to MiniMax-Video, an older model in the same family.

The dissent is worth carrying too. On Hacker News, echelon, who says they have spent "thousands upon thousands of dollars generating video," argues that "nothing touches Seedance 2.0 / 2.5. Not even close," and calls the rankings "whack." coder543 replies that thousands of A/B votes already put H3 ahead of Seedance 2.0. Preference Elo and a professional's eye are measuring different things, and this is the kind of disagreement that usually resolves in about six weeks.

The price claim holds only against the tiers MiniMax chose not to name

The pay-as-you-go page prices H3 at $0.13 per second at 2K and $0.08 at 768P, with the regeneration step at $0.05 and Context-IR billed per token.

That produces a genuinely useful tactic, straight from the price table. Generating at 768P and then regenerating costs $0.08 plus $0.05, which is the same $0.13 as going direct to 2K, except you see the draft first and only pay the top-up on takes you keep.

The launch claim was that "at 2K, H3's per-second price is less than a third of mainstream models." MiniMax names no comparison set. Against live first-party pricing, the claim holds against Veo 3.1 Standard at $0.40 and Sora 2 Pro at $0.30, and fails against several tiers those same vendors sell. Veo 3.1 Fast is $0.12 at 1080p and Veo 3.1 Lite is $0.08. H3 is more expensive than both. It is also more expensive than Sora 2 at $0.10 for 720p. Resolution is not being held constant in any of this, and no independent cost-per-accepted-clip study exists, so the comparison is list-price arithmetic rather than a measured benchmark.

One wrinkle for anyone building a comparison set right now: OpenAI's own deprecations page has the Videos API and the Sora 2 models shutting down on 2026-09-24, with the recommended-replacement column empty.

Resellers all charge above MiniMax's own list. fal is exactly double at both tiers. Atlas Cloud, Segmind and WaveSpeed sit between. Replicate and RunPod do not offer H3 at all. There is a timing detail worth reading as evidence: fal, Segmind, Atlas Cloud and WaveSpeed all published H3 endpoints on 2026-07-31, three days before any weights existed, so the hosted market is almost certainly reselling MiniMax's API rather than serving downloaded weights. That is an inference from the dates, not something any of them state.

Running it locally, with the conditions the headline drops

Start with disk, because the number in circulation is attached to the wrong thing. "42.5 GB" describes a four-file quantized set in the third-party Comfy-Org repack, not MiniMax's release. Cloning MiniMax's own repository pulls 498.47 GB across 280 files, because the two task-family directories duplicate the encoder and VAEs. One task family in diffusers format is about 144 GB. MiniMax's reference deployment command asks for four GPUs.

Now the claim everyone repeated. The Comfy Org launch post says the model can "run locally on a GPU like the RTX 3060." That post does not say 12 GB, does not mention system RAM, does not mention an SSD, and gives no generation time. Every conditioned version of the number traces to a single Discord message from comfyanonymous, the ComfyUI lead developer, relayed to Reddit and quoted by InstaSD:

it's definitively going to be usable on a 3060, 832x480 124 frames takes less than 10 minutes right now on a 3060 + 32GB ram + nvme ssd using 8 bit weights

Seven conditions travel with that: 832x480, 124 frames, 8-bit weights, 20 steps, 32 GB of system RAM, an NVMe drive, and just under ten minutes for five seconds of video. As of 2026-08-04, nobody has publicly reproduced it. The one apparently independent confirmation is a restatement with drifted numbers.

Here is what people with named hardware have actually reported.

No two consumer reports share a full parameter set, so reading across the rows is unsound. Steps are missing from most of them, and steps dominate runtime.

The most instructive row is the pair in the middle. A 4080 and a 5070 Ti, both 16 GB, running the same reference-to-video workflow, came in at 690 seconds and 300 seconds. Another user gave the explanation on the same issue: the 5070 Ti has GDDR7 and PCIe 5.0, so once weights are streaming from system memory, bandwidth decides. Capacity gets you to the starting line. Bandwidth and NVMe throughput decide what happens next. That is a 2.3x spread on nominally equivalent cards, and it should make anyone spec'ing hardware from a VRAM number nervous. Note that only the 4080 figure was self-run; the 5070 Ti number came from a YouTube video the reporter had not verified.

On system RAM, the widely repeated 64 GB floor does not hold up. The only first-party configuration is 32 GB with an NVMe drive. The evidence for 64 GB is a single anonymous report, and it landed on the same day ComfyUI shipped a regression in dynamic VRAM offloading that produced out-of-memory errors on machines with 14 GB free, fixed by passing --disable-dynamic-vram. The feature the entire 12 GB story depends on was broken for some users on the day the weights arrived.

That regression is one of more than ten H3 issues filed against ComfyUI within 48 hours. Others include pure noise output with SageAttention enabled, device mismatches when offloading the VAE on 8 GB cards, and NaN values on FP16 with older datacenter GPUs. This is normal for day one and it is worth knowing before you plan a sprint around it.

Two more things that will not be obvious from a spec sheet. Sampler optimizations underdeliver here, because so much of the wall clock is not sampling: the RunningHub plugin reports a 2.46x faster denoise translating to only 1.35x end to end, with roughly 300 seconds of a 406-second run spent loading, offloading and decoding. And on the same 3060, one commenter relayed by InstaSD puts LTX 2.3 at 720x1280 in about three minutes. If throughput is the constraint, H3 is not the answer this week. Its case is quality, audio and reference control.

The editing capability is real, and it has no strength slider

Reference-to-video is genuinely broad. ComfyUI's documentation describes it as locking in "a character, style, motion, camera move, or voice from reference images, videos, and audio," and it accepts up to nine images, three video clips and three audio clips, twelve files in total, with stereo audio generated in the same pass rather than dubbed on afterwards. For reference-driven work with dialogue, there is currently nothing else open that does this.

The limit is structural. Reference-to-video treats a source clip as semantic context, so it may recompose the frame, reorder cuts or alter timing. There is no denoising-strength control, no pixel-faithful video-to-video, and no pose conditioning. You cannot ask it to change one thing and hold everything else steady.

That is the distance between an editing leaderboard and a production pipeline. Winning a preference pool for "video editing" means people preferred the regenerated result. Shipping a client revision means the approved shot survives the edit, and H3 offers no mechanism to guarantee that.

Prompt adherence has gaps too. One user testing ComfyUI's own demo prompts found a specified whip-pan transition simply replaced with a cut, and reported that other lines in the same prompts "are just totally ignored."

The licence is the part that decides this for most readers

Definition 5 of the MiniMax H3 Community License Agreement reads: "'Excluded Territories' means the European Union, the United Kingdom, the Republic of Korea and the United States of America." Definition 3 makes the Applicable Territory everywhere else, and the grant in Section II attaches only there. Those four are where the licence does not reach.

Section V.4 is the clause most summaries miss:

You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory.

Outputs. The weights, any fine-tune, and the video files themselves. On the plain text, a team inside the Applicable Territory generating clips for a US client is outside the grant. What the agreement never does, across all 84 lines, is say how territory is determined. There is no residency test, no incorporation test, no deployment-location test. The verbs are territorial, so it appears to attach to where the act happens, but that is reading the text rather than anything MiniMax has confirmed.

Two other terms are widely misreported. The USD 20 million figure is not a free-commercial-use ceiling; Section IV.1 is a prior-written-authorization gate, and it measures the revenue of "your commercial products and services" rather than the licensee's total revenue, which on the plain text puts a large company with a small product line below the line. And the mandatory requirement to display "MiniMax H3" in a commercial interface sits in Section IV.2, while the trademark permission in Section VI.2 is scoped to the encouraged notice in Section III.3. The licence compels a display it does not quite license.

Then the detail that has not appeared anywhere in the coverage. Definition 15 says the licensor is Nanonoble Pte. Ltd. Nanonoble Pte. Ltd. is also a named defendant in case 2:25-cv-08768 in the Central District of California, filed on 2025-09-16 by twelve Disney, Universal and Warner Bros. Discovery entities, where the motion to dismiss was denied on 2026-05-22. The entity being sued in California is the entity writing the exclusion that covers the United States.

The licence FAQ gestures at this, listing "ongoing copyright-related legal proceedings specifically concerning generative video AI" as one of three reasons for the US position, alongside the EU AI Act and regulatory uncertainty in the UK and Korea. A MiniMax staff member answering the licence thread on the model's Hugging Face page was more direct: "This regional carve-out stems from our ongoing generative video copyright litigation with major Hollywood studios." Three conditions on that quote. It covers the United States only, and the same person gave regulatory uncertainty as the reason for the other three territories. It is a staff comment in a user-opened thread rather than a corporate statement. And a denied motion to dismiss is a pleading-stage ruling, not a finding of liability.

The asymmetry MiniMax states openly: the API is globally available, the weights are not. Excluded-territory organizations can apply for a bespoke licence through a form, and one user asked in that thread why the restrictions were not communicated before a release marketed on openness. The reply linked the form.

Whether a third-party host changes your position is unresolved. Atlas Cloud says a hosted API is a different legal relationship that makes the territorial restriction inapplicable to its customers, which is a vendor asserting the legal effect of its own product on a question MiniMax has answered only about its own API. This article reports licence terms; the question of which territory your organization acts in belongs with counsel.

Where this leaves the decision

If you are in the four excluded territories and you want output, the API is the whole offer, and it is a good one. Nothing about the weights changes your week.

If you are in those territories and you want weights, the comparison is not H3 against Wan2.2 or Seedance. It is H3 against LTX-2.3, which also generates synchronized audio and video from a single model, which is roughly three times faster at higher resolution on the same card, and whose licence carries a revenue trigger but no territorial exclusion. LTX-2.3 is usable today. H3 is not, whatever the Elo says.

If you are outside them, doing reference-driven work where dialogue and character consistency matter, this is the strongest open artifact currently available, and you should budget your first week for offload configuration rather than for output.

The interesting thing about H3 is not that MiniMax kept part of it back. It is that the architecture made keeping it back cheap. A super-resolution network is a file, and files get released. A pass that re-reads the original context through the base model is a pipeline, and a pipeline is a service. The more elegant the design got, the less of it had to ship.

References

Primary sources

1. MiniMax. MiniMax H3 model card. Hugging Face, 2026-08-03.
2. MiniMax. MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities. MiniMax Research, 2026-07-31.
3. Nanonoble Pte. Ltd. MiniMax H3 Community License Agreement. 2026-08-02.
4. MiniMax. Q&A About License. Hugging Face, 2026-08-03.
5. MiniMax. Pay as You Go pricing. MiniMax Open Platform, accessed 2026-08-04.

Measurement

1. Artificial Analysis. Video editing leaderboard and text to video leaderboard. Read 2026-08-04.
2. SGLang. MiniMax-H3 deployment cookbook. Accessed 2026-08-04.

Implementation

1. Hugging Face. MiniMax-H3 diffusers pipeline. Branch minimax-h3, accessed 2026-08-04.
2. Comfy Org. MiniMax H3 Day-0 Support in ComfyUI. 2026-08-03.
3. Comfy Org. MiniMax H3 tutorial. Accessed 2026-08-04.
4. Comfy-Org. Why is a 5070 Ti faster than a 4080. GitHub, 2026-08-03.

Litigation

1. Disney Enterprises, Inc. et al v. Minimax et al, case 2:25-cv-08768. US District Court, Central District of California, filed 2025-09-16.
2. Loeb & Loeb LLP. Disney Enterprises, Inc. v. Minimax. Case note, 2026-06.

Commentary and practitioner reports

1. Hacker News. Discussion of the ComfyUI day-0 post. 2026-08-03.
2. InstaSD. MiniMax H3 Open Weights: What We Know Before the Drop. 2026-08-01.
3. Kingy AI. MiniMax H3 Review: Benchmarks, Specs and Hardware. 2026-08-03.
4. Invide Labs. MiniMax H3 Open Weights Release Comes With Major Limits. 2026-08-04.