IT EN
Start a trial

Operational notes Observatory

Is MiniMax H3 open weight? The weights are not out yet

6 min read

A red theatre curtain, closed, photographed head-on, in black and white
The curtain is announced as open. It has not risen yet: that is where H3’s weights stand today.

On 31 July 2026 MiniMax published a blog post announcing H3, a multimodal model — text, images, video, audio — presented as the moment its video line moves to open weights. The same day, fal.ai announced it was the “official API partner” for H3, and generation is already usable online through that channel. Anyone looking for weights to download, though, finds a single sentence in the announcement itself: “we plan to open up the model weights in the coming days, subject to applicable laws and regulations.” That is not a release. It is a stated intention, with a legal condition attached and no date. Before anyone writes “open weight” into a tender specification, it is worth checking what actually exists today, a day after the announcement.

What is declared, and what is missing

The specifications MiniMax publishes are detailed and technical: resolution up to 2K (default) or 768p, clips up to 15 seconds, native stereo audio, four named components — H3-Contextual Omni Representation, the H3-VAE tokeniser (a declared “4x” gain in effective sequence length), the H3-Omni Transformer (roughly 30% more training throughput, without stating the baseline), and H3-In-Context Regeneration for 2K upscaling. On price, MiniMax states that the per-second cost at 2K is “less than a third” of “mainstream models” and at 768p “less than half” the price of their 720p — without naming the competitors or the measurement method: that is a commercial claim, not a benchmark a third party can check.

What the announcement does not contain matters just as much: no parameter count, no separate model card, no report with stated evaluation conditions. It is not possible today to work out the bytes needed to run it locally — the exercise we ran for Inkling-Small or for Soofi S — because no weight file exists yet to measure. Anyone pitching you H3 today as an on-premise component is describing a product that, as of this article, cannot be installed anywhere: it is reachable only through an API, on terms MiniMax can still change before the day it actually ships.

The licence in circulation is not the one behind the previous model

A secondary source — the aggregator nationpress.com, citing the independent benchmarking firm Artificial Analysis, with no direct link to a MiniMax statement or a published licence text — reports that H3’s weights will ship under a “MiniMax Community License”: free commercial use for organisations under US$20 million in annual revenue, and “separate licensing terms” above that threshold. We could not verify this against a primary text, because that text — if it exists — is not public yet. We treat it as unconfirmed, not as fact.

It is worth comparing this with what MiniMax has already published for its own language model, M2: the LICENSE file in the MiniMax-AI/MiniMax-M2 GitHub repository is a modified MIT licence, and the modification requires commercial products above 100 million monthly active users or US$30 million in annual revenue to display “MiniMax M2” prominently in the interface. That is an attribution threshold, not a ban: above or below that number, commercial use stays free — only the requirement to credit the source in the interface changes. If the threshold reported for H3 turns out to match what has circulated — a revenue ceiling above which commercial use itself needs a different agreement, not just attribution — that would put it structurally closer to Kimi K3’s licence than to MiniMax’s other models. It is exactly the kind of detail an open-weight licence has to be read in full to catch — and, for H3, there is nothing yet to read.

The problem no weights licence can settle

There is a second layer of risk that the weights’ licence — whatever it turns out to be — does not touch by definition: what the model generates. On 16 September 2025 Disney, Warner Bros. Discovery and NBCUniversal (Comcast) sued MiniMax in the US District Court for the Central District of California (Disney Enterprises, Inc. v. Minimax, case 2:25-cv-08768), alleging that MiniMax’s Hailuo video generation service — the proprietary product H3 replaces — produced, on request, images and videos featuring copyrighted characters, including Spider-Man, Darth Vader and Shrek, without authorisation. On 26 May 2026, Judge Stanley Blumenfeld denied MiniMax’s motion to dismiss the case, letting the claims over unauthorised character generation go forward. On 31 July 2026, when H3 was announced, the case was still open.

A weights licence governs what you may do with the model. It says nothing about what the model produces when someone asks it for a protected character, and that is the part of the Disney case that concerns use, not training. An organisation that deployed H3 to generate content — marketing, training material, institutional communication — would inherit the same problem MiniMax is already fighting in court, regardless of how permissive the weights licence turns out to be: whoever generates the output answers for the output, and no software licence clause absolves a copyright infringement committed by producing it.

What to do, in practice

  • Do not put H3 into a tender as an available component: today it is an API service, not a package of weights. Wait for the actual release before writing “on-premise” next to its name.
  • When the weights do ship, get the licence text from the source — MiniMax’s own Hugging Face or GitHub repository — not a citation from an aggregator citing another source: the US$20 million threshold is, today, unverified against a primary document.
  • Treat the weights licence and the output risk as two separate assessments. Procurement handles the first. The second needs a technical control that blocks the generation of recognisable characters, brands or logos, plus a contractual indemnity clause if you use the model for externally facing content.
  • If the use case is for a public body or institutional communication, the second assessment outweighs the first: synthetic content that can survive a lawsuit is not the same thing as cheap content.

How we solve it

A video-generation model announced as open but not yet downloadable is a candidate to watch, not to install. When — and if — the weights ship, the question that matters does not change: where the model runs, who sees the outputs before they go out, and who answers if an output reproduces something it should not have. That is guaranteed by the environment the system runs in and the controls you put around it, not by the name on the licence. That is why we deliver in two modes: on-premise, inside the client’s own infrastructure, or on a dedicated CSIDIA cloud — reserved for a single client, reached over a dedicated VPN, with a data centre resident in Italy and premises staffed directly by us. In both, the model stays a replaceable component, with a filter on outputs and a log of what was generated and who approved it, before anything crosses the perimeter. The fuller argument on weights, hardware and vendor dependency is in our page on hardware and open-weight models; the method we use to build the system around a model, whichever one it is, is set out on our platform page.

Are you evaluating a video-generation or multimodal model for corporate or institutional content? Let’s talk it through in a thirty-minute session: we separate the weights licence from the output risk before they become a production problem.

Sources