On August 31, MiniMax surged by 16.18%. From its July low, the stock has rebounded by 92%, with a market capitalization returning to 124.9 billion Hong Kong dollars.
Just three months ago, the market had held a funeral for this company. The stock had plummeted by over 80% from its peak, causing a 300 billion Hong Kong dollar market capitalization evaporation.
On August 26, MiniMax released its interim report, showing a 283% year-on-year revenue growth, with a 703% growth in B2B revenue. By August, the company's ARR had exceeded $800 million, and the token consumption in July was 20 times that of January.
What brought MiniMax back into the spotlight a few days later was a model it didn't even create itself.
H3 Max.
Infinite Channel
On August 3, MiniMax open-sourced H3, a 33 billion-parameter multimodal model capable of processing text, images, videos, and speech simultaneously. fal took its open-source weights, continued training, further optimized them with its own inference infrastructure, and eventually created H3 Max.
According to fal, when generating a 5-second video clip, H3 Max's throughput can reach approximately 35 times that of MiniMax's official endpoint.
On August 29, Rehan Sheikh, an engineer at fal, a U.S.-based generative AI infrastructure company, connected H3 Max to Twitch to start a live stream. The content of the stream was not pre-prepared videos. As the video played, the model generated content in real-time. Each 5-second segment at 768p resolution with synchronous audio was generated in approximately 3 seconds.
Generation was faster than playback.
Due to this feature, viewers could influence the storyline through the barrage of comments. Within seconds, a new plot twist would appear on the screen. As long as the machine kept running, this channel theoretically could continue indefinitely. Over a weekend, it attracted around 5 million views. Some who disliked AI-generated content even gave it a nickname, calling it the "Infinite Garbage Disposal Machine."
What sets them apart is not the underlying model, but rather the post-training, inference optimization, hardware scheduling, and deployment methods after the model is delivered.
One media outlet directly titled their coverage:
《Post-Train Beats Parent》
Post-Training Beats the Parent Model
While this may seem like a beautiful engineering optimization, it encountered a larger issue behind the MiniMax backlash.
If a model can be open-sourced, taken away, further trained by others, and even outperformed, then what should the market value from a model company?
To answer this question, let's first take a look at the company, fal.
This company does not train base models. Its business involves taking models trained by other companies, running them on its infrastructure, and then selling them to developers. The faster a new model comes out, the faster its shelves are updated.
Once the models are onboarded, fal re-optimizes for inference, hardware scheduling, and deployment to drive speed and cost to a level more suitable for commercial use. Over 2.5 million developers use its generative models through fal, and the company's ARR is around $400 million, with a team of only 70 people.
In the past year, fal's revenue has grown approximately 60-fold, and its valuation has reportedly risen from $4.5 billion to around $8 billion.
Redpoint conducted an in-depth interview with the fal founding team, where Pat Grady, a partner at Redpoint Ventures, mentioned that in the second and third quarters of 2025, the half-life of the top five video models was only 30 days.
The term "half-life" originates from physics, originally referring to the time needed for half the atoms in a radioactive substance to undergo decay. Here, it signifies the time it takes for approximately half the models on the Top 5 Video Models list to be replaced by new models.
A model that has just taken the top spot may be replaced by another one in a month. Regardless of model changes, developers always need a platform to run them.
fal holds this position.
Developers pay per call for each image or video generated. By integrating with platforms like fal, they can switch to the best model at any time without reconfiguring APIs or building infrastructure from scratch. Models may change every month, but the platform's position is hard to replace.
The CEO of fal, Burkay Gur, once said, "When content generation becomes infinite, limited things become more valuable."
This company itself is a business born after the rapid depreciation of the model. H3 was done by MiniMax, but the users, revenue, and invocation data of H3 Max will initially remain in fal. The more successful H3 Max is, the sharper this issue becomes.
Moreover, fal is a private company, not available for purchase on the open market.
From HK$1100 to HK$200
When MiniMax went public, the market believed in another story.
This company does text, voice, and video models simultaneously, with Talkie and Sea Snail AI on the consumer end and an API on the business end. Compared to start-ups that only focus on a single model, it appeared more like a complete multimodal business loop.
On January 9th this year, MiniMax went public at an IPO price of HK$165, with a public subscription oversubscribed by 1848 times, and the first-day closing price surged by 109%. On March 10th, the stock price intraday broke through HK$1100, surpassing Baidu in market value.
That was when the market most believed in this story.
A turning point came in June. On June 1st, MiniMax released the flagship model M3, with 428 billion parameters. Compared to GLM-5.2 and DeepSeek V4 at the same time, its model size was smaller. The official emphasis was on training and engineering efficiency, with a system that could run continuously for 12 hours without human intervention, autonomously train four models, and accelerate the CUDA kernel by 9.4 times.
Then the problem arose. The community found that some of the officially announced results came from self-testing, and there was a significant gap between third-party actual tests and public numbers. Simultaneous adjustment of billing rules also significantly increased the rate of consumption of credits for some old users.
On June 5th, the company publicly apologized, stating, "M3 requires more computational power, which is our oversight." Ten days later, the flagship model, which had only been online for two weeks, was permanently discounted by 50%; subsequently, on June 16th, Zhigu released GLM-5.2, taking the top spot on the open source leaderboard, and the lead of M3 lasted only two weeks. On the first day of lifting restrictions on July 9th, MiniMax fell by 20%, and a few days later, the stock price had fallen by over 80% from its peak.
Just one day after the market crash, Yan Junjie announced that he would no longer receive a salary until AGI is achieved, while also allocating his personal share to reward the team and support open source. On that same day, the company completed a HK$16 billion private placement, which was oversubscribed by about 7 times.
By August 26, MiniMax released a rather impressive interim report, showing a 283% year-on-year revenue growth, a 703% growth in B-side revenue, and continued rapid growth in research and development investment. However, the next day, the stock price only rose by 4%.
It is evident that the market is still waiting for something else.
Core Assets
The biggest trouble for model companies is that product updates are happening too quickly.
In early June, M3 just took the top spot, only to be surpassed by GLM-5.2 half a month later; followed by releases of DeepSeek V4, GLM-5.3, and Kimi K3 in succession. The current leader may easily be dethroned by the next model, as the window of being ahead is only a few weeks.
Being in the lead can only be counted by weeks. Clearly, relying solely on one model cannot support a company's long-term valuation.
So, over the past year, the market has been searching for the model companies' true core assets.
The most direct answer is the next-generation models, but the next generation will also quickly become outdated. As long as the industry continues to iterate at a high speed, the act of "creating a new model" can only ensure that a company is not eliminated but does not explain why it is worth holding onto in the long term.
A more convincing answer is the ability to create models.
Many model releases this year have proven that with the same base, relying solely on later training and environment can lead to significant leaps in capability.
Both Spectrum GLM-5.3 and GLM-5.2 use the same 753B parameter base, and what truly changes is the larger-scale training environment and reinforcement learning post-training. The Terminal-Bench 3.0 score jumped from 4.6 to 28.3 without retraining the base, resulting in a significant improvement.
Later, Tang Jie referred to this as a "controlled variable experiment."
During training and evaluation, Kimi K3 established over 50 million sandbox environments, with a significant amount of machine time spent waiting for model inferences. Operations such as freezing, restoring, and forking were compressed to tens of milliseconds to allow the model to repeatedly attempt, fail, and retry in a massive real-world environment.
Cline even allowed K3 to modify its own training scaffold. After 17 hours with no human intervention, the score increased from 77.5 to 88.8.
The truly difficult-to-replicate part for model companies is transitioning from a single set of weights to a full system of production weights, fine-tuning weights, and evaluation weights.
Anthropic spent over $1 billion in one year on reinforcement learning environments. OpenAI purchased tens of thousands of Macs for its computer-use agents for the same reason.
The barrier to entry and investment required to build this system are rapidly increasing, and it is not solely in the hands of model companies.
Cursor's Composer 2 is a prime example. While it is based on the open-source Kimi K2.5 from the Dark Side of the Moon, the market value of Cursor is clearly unrelated to Kimi K2.5.
Cursor has accumulated a large amount of real programming behaviors, developer workflows, and a reinforcement learning system trained on this data. The foundational model comes from elsewhere, but what truly matters is what is left behind after the model is used.
fal follows a similar path. It takes the H3 weights and creates the H3 Max. Users generate videos on fal, revenue flows into fal first, and usage data stays in fal.
The upstream model companies provide the most expensive raw materials, but downstream, it is possible to accumulate the truly compounding assets. Technological complexity and value-capture ability do not always directly align.
Thirty years ago, a similar situation occurred in the personal computer industry. The chip was the most complex part of the machine, but consumers recognized Dell and Compaq. Intel later spent over a decade reintroducing itself into consumer consciousness through "Intel Inside."
Model Minus Model Company
Therefore, to assess the value of a model company, one might consider a subtraction. After deducting the model weights, replicable inference optimization, and the post-training capability that can be taken by third parties, what remains is truly what this company can hold onto in the long term.
MiniMax happened to conduct such an experiment proactively. On August 3, it open-sourced H3. Within a short period, hundreds of derivative models emerged, with downloads reaching tens of millions. Chip platforms such as Huawei Ascend, Muxi, AMD quickly completed adaptation, and numerous developers and partners joined the ecosystem.
While the weights were open-sourced, enterprise contracts and API revenue did not diminish—just in July before the open-sourcing of H3, MiniMax's token consumption was already 20 times that of January, and this was still on its books.
More importantly, H3 did not release everything. The H3-Context-IR and Regenerate-2K two hosted modules are still controlled by MiniMax. To output higher-resolution 2K videos, you still need to call the official service. This is a typical open-source strategy, where the underlying capabilities are diffused, allowing the ecosystem to grow, while retaining the parts that can truly generate commercial returns on the server side.
Therefore, the significance of H3 Max to MiniMax is not just "someone has optimized my model faster," it demonstrated for the first time that subtraction.
The mid-year report provided some answers, with B-side revenue growing by 703%, ARR exceeding $800 million, showing that open source has not yet caused all value to be lost. However, H3 Max also demonstrated the other half of the risk, that user relationships, call revenue, and usage data may settle on downstream platforms.
Long vs Short
In late August, MiniMax's stock price rebounded close to double from its low point, while the short ratio simultaneously rose to about 20%, reaching a high point.
What the long side sees is what remains after subtraction. For example, enterprise customers, API revenue, actual calls, data feedback, and the system's ability to continue training the next generation of models. In the public market, such assets are scarce, there are few model companies available for purchase, hence MiniMax carries an obvious scarcity premium.
The popularity of H3 Max has given a very intuitive reference to this money. If a company like fal is worth $8 billion, and one of fal's hottest new products is built on H3, then how much is the company behind H3 worth?
This is the long side's perspective.
The short side's perspective is, if the weights can be taken for free, post-training can be done by fal, and the inference speed can be improved by a third party by 35 times, then why should all these capabilities still remain in MiniMax's valuation?
This also explains MiniMax's abnormal performance in the past few months. Financial reports can only reflect how much money the company is making now, but what the market really cares about is what it can retain after open-sourcing. H3 Max has, for the first time, put this answer on the table.
Feedback Loop
After releasing H3, MiniMax is actually facing a very old business problem. A company can voluntarily give up one layer of scarcity, provided it knows where the next layer of money is.
Red Hat has made such a choice before.
The source code of Linux has never belonged to Red Hat; anyone can download, modify, and redistribute it for free. What Red Hat eventually sold was not Linux itself, but a stable version, long-term support, technical support, and certification system that enterprises were willing to pay for.
The wider the adoption of open-source software, the larger the market for this suite of services. In 2019, IBM acquired Red Hat for $34 billion, clearly not for a copy of Linux source code that anyone can download.
Anything released must be able to bring back value.
This is also what MiniMax truly needs to prove after open-sourcing H3.
Now the market has seen diffusion. The model has been downloaded, adapted, retrained, and third-party products like H3 Max have begun to emerge.
What we need to observe next is whether this diffusion can ultimately be reflected in MiniMax's own revenue and assets.
It could be more enterprise customers, continuous growth in API consumption, or the data and feedback left by developers during usage, which will eventually re-enter for further training and become part of the next generation model.
If these elements can form a cycle, the more H3 is taken away, the more entry points MiniMax will receive.
If the cycle does not form, the situation will be completely different.
The model's influence may continue to grow, the ecosystem may become more prosperous, but the most valuable customer relationships, user data, and transaction entry points may ultimately settle in others' hands. By that time, MiniMax has provided the underlying capability but has not captured most of the value generated by this round of diffusion.
Therefore, what H3's open-source initiative truly needs to test is not the download volume or how many companies announce compatibility. Those numbers can only prove how widely it has spread.
What needs to be observed in the next few quarters is how much money can come back after the release. Red Hat has enterprise service contracts, Google has search and advertising. MiniMax also needs to find its own pipeline.
----END-
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
