MiniMax Open-Sourced H3 to Buy Distribution. The Market Paid, Then Took It Back

MiniMax gave away the weights for H3, the most capable video model it had built, in the first days of August. Nine weeks later, the company that published them closed at HK$303 a share, 77% below where the stock peaked in March.

That looks like a contradiction, and MiniMax explained why it did it anyway. "Closed-source models have long dominated video generation, with slower iteration and a less open ecosystem than fields like large language models," it wrote in the launch post that put H3 online on July 31, 2026. Opening a frontier model is the opposite of the obvious move: you pay for the compute, then hand the artifact to everyone.

The Case for Giving H3 Away

H3 is a general-purpose model rather than a video generator with extras bolted on: it reads text, images, video and audio in one context and generates up to 15 seconds at 2K with native stereo sound, with reference-to-video and editing endpoints on top. What MiniMax open-sourced is the 768p half of that. The release is H3-Base, a 33.1-billion-parameter omni transformer, published as two task checkpoints, FL2VA and Ref2VA, under the MiniMax H3 Community Licence, dated August 2, 2026 and royalty-free for commercial use.

The two modules that make the 2K pipeline work did not go with it. H3-Context-IR, the preprocessing stage that rewrites your prompt into what the model consumes best, and H3-Regenerate-2K, which rebuilds a 768p result at 2K, are API-only. The free model is deliberately the slower, lower-resolution half.

Then there is the licence itself, which reads like a distribution plan in legal language. It excludes the European Union, the United Kingdom, the United States and South Korea, the four markets with the most money in them, and forbids using the weights or the output to train other models. Host H3 for other people and you must display "MiniMax H3" in your interface and run content safeguards. We walked through those terms when they landed; the short version is that the weights travel, the brand travels with them, and a creator in an excluded territory has one obvious route to commercial use, which is MiniMax's own API.

For a giveaway it has travelled: 4,299,737 downloads in the past month on MiniMax's official Hugging Face repository and 20,497,092 on ComfyUI's repackaged copy as of September 19, which is 24.8 million in about seven weeks.

The Fork That Beat the Base Model

The most instructive thing to happen since is what fal did with those weights. fal was a day-0 ecosystem partner and hosts the whole H3 family, including the 2K, reference and editing endpoints, which on fal's stack run about 15 times faster than MiniMax's own inference. But fal also post-trained its own version on top of the open base, adding new training data and spending a large share of its compute on verifiable reinforcement-learning tasks,

H3 Max now leads the boards buyers actually read: first on Design Arena's image-to-video leaderboard at an Elo of 1,341, above the base model it was trained from, and first on Artificial Analysis's image-to-video-with-audio board at 1,201, with a margin of error near 11 points across 2,177 samples. The speed claims are louder still - fal measures 2.46 seconds of generation for a 5-second 768p clip on H3 Max and 1.54 seconds on Turbo.

The price moved the other way. Both models launched at 75% off until September 14, 2026, so that 5-second 768p clip cost $0.10; from September 15 it lists at $0.08 per second, or $0.40 for the same clip. Even at list, 15 seconds at 768p is $1.20 on H3 Max against fal's published prices of $1.50 for Wan 3.0 at 720p, $1.89 for Kling v3 Standard with audio and $7.10 for Seedance 2.5. That is the structural point worth keeping: open weights turn a model into a substrate, and a serving company with a good inference stack can out-tune the lab that invented it. MiniMax's endpoint is no longer the only way to get H3, nor the best measured version of it.

What the Share Price Actually Says

MiniMax listed in Hong Kong on January 9, 2026, with the prospectus capping the offer price at HK$165.00 a share. Its first session closed at HK$345.00, then it ran to HK$1,330.00 on March 18 and fell almost all the way back, to an intraday HK$186.20 on July 27, closing that day at HK$229.60. Every figure here comes from the daily price history for 0100.HK.

The open-source week lands in the middle of the recovery. The launch post came on July 31, the licence is dated August 2, and the weights were public that week: the stock closed at HK$247.20 on August 3, up 7.2% from July 31, and at HK$326.40 on August 7, an advance of 32% in four sessions, gaining 40.4% from the July 27 close to August 10. It reached HK$355.00 by September 3, then slid to HK$303.00 at the last close on September 18 - 14.6% below that September peak and 77.2% below March.

The honest caveat matters more than the correlation. The rebound began before the weights were public, Zhipu had listed the day before MiniMax, and the whole AI sector was moving. Nothing here proves that open-sourcing H3 caused a 40% rally; what it shows is that the giveaway did not scare the market, and that for a fortnight the market treated it as good news before handing most of it back.

Three things follow. Adoption is not revenue: 24.8 million downloads can coexist with a share price 77% off its peak. The licence is the business model, because excluded territories and hosted-service branding turn an open release into a funnel rather than a gift. And the crown stays closed, because 2K regeneration and context preprocessing remain API-only, making the free version a different product rather than the same one at no cost.

The signals to watch: whether downloads turn into paid API traffic, whether Regenerate-2K and Context-IR ever open, whether creators in the four excluded regions get commercial terms, and what fal post-trains next now that it has a template for beating a partner's base model.

The weights are free, the top of the pipeline is not, and the market has already priced the difference. What H3 bought MiniMax was not a valuation but a position: the next generation of builders learned this one on MiniMax's terms.