Why is the secondary market not buying in? Since the start of September, MiniMax has been making frequent moves in overseas markets. On September 16, three of MiniMax's products were selected as the only domestic large model representatives included in an AI learning program launched by SkillsFuture Singapore, standing alongside ChatGPT Plus, Google AI Pro, and Microsoft 365 Personal as available services.
On September 3, HUMAIN, an AI company under Saudi Arabia's Public Investment Fund, released Humain-M3, which is built on the MiniMax M3 foundation and continued pre-training with over one trillion tokens of native Arabic language data. It ranked first in average scores across seven Arabic benchmark tests. Driven by this, MiniMax's stock price rose for two consecutive days on September 17 and 18, with Hong Kong shares closing up nearly 19% on the 18th. As of the close on September 21, MiniMax's stock price stood at HK$298 per share.
Compared to its peak stock price of HK$1,330 per share, outsiders may now be more concerned about one question: when will M3.1 and M3 Pro be released? During the earnings call on August 26, MiniMax founder and CEO Yan Junjie stated, "With the continuous improvement of both the M-series and H-series pipelines, the model release cycle will be significantly shortened. We can look forward to new products such as M3.1, M3 Pro, and H3.1. The parameter scale of M3 Pro is expected to increase to approximately 3 trillion, equipped with the MSA 2.0 architecture, and computing efficiency is expected to increase by approximately three times."
However, as of now, there has been no news of M3.1's release. Meanwhile, the iteration pace of domestic large models has visibly accelerated in the second half of 2026—Zhipu released GLM-5.3 on August 14, Kimi released K3 with 2.8 trillion parameters on July 17, and DeepSeek V4 Pro also launched its official version on August 13. Although MiniMax released the multimodal video model H3 on July 31, under the current evaluation system, text model capabilities receive more attention. M3, due to its rushed release, insufficient preparation, and timing that placed it awkwardly between GLM-5 and K3—both of which excel in areas like Coding and Agent—coupled with strong developer backlash over its pricing strategy, appears particularly awkward in the competitive landscape.
From a product iteration logic perspective, covering up a previous misstep with a superior model would normally be the optimal solution, but the company's prolonged failure to release M3.1 and M3 Pro has also led to external doubts about MiniMax's model capabilities. Currently, when people mention Zhipu, the capital market thinks of Coding; when they mention Kimi, they think of Agent clusters; when they mention DeepSeek, they think of a focus on text and price advantages. These labels and partial leadership in model capabilities have indeed given these companies clear valuation anchors in the capital market and made user perception of them more distinct.
However, MiniMax's multimodal strategy has not yet formed an absolute leading advantage in any single capability or price point. The outside world perceives its positioning as perpetually ambiguous, and its valuation system consequently wavers. In June 2026, JPMorgan downgraded MiniMax from "Overweight" to "Neutral" and significantly cut its target price, with the core reason being that MiniMax's flagship model M3 lacks pricing power and its model capabilities are still in the catch-up phase. This means that going forward, MiniMax must produce a sufficiently powerful model to prove that its chosen path is correct.
In reality, the questions MiniMax needs to answer go far beyond the model itself. From to C to to B On August 26, MiniMax released its first interim results after listing. The company's total first-half revenue was $117 million, a year-on-year increase of 283.1%, reaching 1.5 times its full-year 2025 revenue in just six months. Q2 revenue grew 81.8% quarter-over-quarter; July Token consumption had already reached 20 times January's level; and August ARR (Annual Recurring Revenue) further improved to over $800 million. Gross profit was $20.81 million, up 464.8% year-on-year, with gross margin improving from 12.1% to 17.9%, a year-on-year increase of 5.8 percentage points.
But another set of figures also exists: the company's first-half R&D expenses were $297 million, up 138.8% year-on-year; the period's net loss was $358 million, narrowing 11% from the $402.2 million in the same period last year; however, after excluding share-based payments, fair value changes of financial liabilities, and listing expenses, the adjusted net loss was $293 million, expanding 111.2% year-on-year. The adjusted net loss margin narrowed from approximately 455.9% to approximately 251.4%, which looks like an improvement, but the main reason is the larger revenue base—the absolute amount of loss is still accelerating.
More noteworthy than the profit and loss figures is the dramatic shift in revenue structure. In the first half of 2025, MiniMax's AI-native products (consumer products like Hailuo AI, Talkie, and Xingye) accounted for approximately 70% of revenue, serving as the company's absolute pillar. By the first half of 2026, B-end open platform and enterprise service revenue surged from $9.2 million in the same period last year to $73.9 million, a year-on-year increase of 703.1%, with its share leaping from 30.3% to 63.4%, becoming the largest revenue source; AI-native product revenue, although up 100.9% year-on-year to $42.6 million, saw its share drop to 36.6%. By August, to B accounted for approximately 80% of ARR, with to C at only about 20%.
In just half a year, a company known to ordinary users for its consumer products completely reversed its revenue structure. However, this shift was not entirely without warning. At least on the Coding front, internal discussions at MiniMax began far earlier than outsiders realized. On October 27, 2025, MiniMax officially open-sourced and launched MiniMax M2, touted as designed specifically for agents and code. Yu Yang, an entrepreneur and former MiniMax employee, told the media: "The company's judgment on Coding actually emerged when Claude first surfaced in 2025. At that time, IO (Yan Junjie) had already shared his views at the CD meeting."
The CD meeting is an internal meeting MiniMax holds every Friday at noon, where Yan Junjie reports business developments to employees, answers their questions, and shares his views on AI frontier developments. Therefore, in Yu Yang's view, MiniMax's commercialization push toward the B-end was foreseeable long ago. "Because entering the Coding track means serving B-end customers." At a MiniMax developer offline exchange event in June 2026, Yan Junjie also mentioned that two years ago, he had asked Liang Wenfeng whether to pursue AI Coding. He said the consensus at the time was that only about one to two million people in China could write code, which didn't seem like a sufficiently broad market, but clearly that has changed now. AI Coding can indeed give more ordinary people productivity.
However, early positioning does not mean cognitive leadership can translate into market advantage. MiniMax's model parameters and capabilities in the Coding track have not formed a sufficiently strong cognitive barrier among the developer community. This cognitive lag quickly reflected in the capital market. When the market's perception of model capabilities failed to match its technical investment, the valuation logic began to waver. The unlock pressure and valuation regression in the secondary market followed. Despite over 80% of pre-IPO and cornerstone shareholders publicly stating they would continue to hold, and two major strategic shareholders, Alibaba and miHoYo, clearly expressing support, the company's stock price still fluctuated violently, falling over 80% from its peak.
After MiniMax's earnings report on August 26, the stock price only rose 3.83% the next day. On August 31, after the live stream built on H3 Max was launched, MiniMax's stock price quickly surged, rising nearly 20% intraday and closing up 16.18%; on September 1, the rally continued, with cumulative gains exceeding 20% over two trading days. This also demonstrates that the market still believes in the narrative of model capability. Yan Junjie also emphasized this point again during the earnings call. When sharing the company's positioning and strategy, he said: "The intelligence improvement driven by large language models has almost no end. From the practical application of Coding and Agent capabilities since the second half of last year, to more autonomous and creative completion of long-horizon tasks, to the foreseeable delivery of reliable end-to-end results, we continue to pursue higher intelligence."
The 'Mishap' of M3 and the 'Turnaround' of H3 On June 1, MiniMax released M3. The model has a total parameter scale of 428B, with 23B active parameters, adopting a Mixture of Experts (MoE) architecture and natively supporting million-token context. Viewed in the competitive landscape at the time, this parameter scale looked somewhat awkward. When M3 was released, domestic models like Qwen3-Max-Thinking, DeepSeek V4 Pro, and Kimi K2.6 had already been deployed, while overseas competitors included GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.7. Trillion-scale parameters were becoming the entry ticket for top players. Two months later, Kimi K3's open-source release with 2.8 trillion parameters and Qwen-3.8 Max reaching 2.4 trillion further widened this gap.
On the eve of its release, expectations for M3 were high, and internally the company hoped model capabilities would reach a new level. But after launch, many developers who tried it said M3's actual performance did not meet expectations. Some insiders also admitted that M3's release was somewhat "hasty," or "rushed." Individual developer Alex said that his current daily driver model is Kimi K3, and when his quota runs out, he tries Zhipu's GLM, with MiniMax as a backup option. "My feeling about MiniMax is that its capabilities aren't that strong, but it does the work reliably," Alex explained his selection logic. He's not a software engineer with strong discernment skills; when asking AI to perform functions, he can't understand most of the questions the model asks him. In such cases, he prefers to hand decision-making authority to the AI, letting it decide how to proceed, then verify the results the AI produces. "So I need a stronger model, not a dumb but fast one, because the latter greatly increases the frequency of me checking its work." Therefore, he prioritizes Kimi or stronger SOTA models.
Since his daily work involves providing software for the industrial sector, Chen Lijun, founder of Yita Industrial Intelligent Technology, has higher requirements for model capability and code accuracy, and M3's performance could not meet his delivery requirements. He believes the root of the problem lies in parameter count. Using Qwen models as an example: "3.7 Max and Plus have the same total parameter count, but Plus is multimodal. Under the same parameters, adding multimodal parameters causes Coding quality to drop sharply. Switching to the trillion-scale 3.8 Max, even as a multimodal model, its capability is on par with or even higher than 3.7." In Chen Lijun's view, within models specifically trained for Coding or among models of the same parameter class, pure text models are generally stronger than multimodal ones. For Chen Lijun, M3's problem is insufficient parameter count.
He calculated: "What really consumes Tokens is rework. I'd rather spend more tokens on upfront research, nail each module in one pass, and then call each other later for maximum benefit. If you generate something and then ask AI to revise it, it might delete earlier parts while modifying—that's very uncontrollable, and it's a deep pit I've personally fallen into." In his view, the model's "cost-effectiveness" doesn't hold up in daily workflows: money saved on low prices may be paid back double in rework. However, M3 isn't necessarily unusable. Developer Xiao Feng emphasized that method matters more than the model itself. In his view, everyone uses models differently; some people may not even use the built-in Skills, and the problem may not lie with the model. "You give the overall direction to AI, and AI helps you implement it, but before execution, you should use Plan to plan first, generate a to-do list, check the contents yourself against requirements, and then have it execute based on the list."
He also mentioned that deleting 80% of Claude Code's prompts can achieve the same effect; the key lies in clear requirement descriptions and structure. "Choose the direction first, then execute; you can't rely entirely on AI." Xiao Feng said he uses MiniMax daily for Coding, images, and image recognition—for example, using DeepSeek to generate text, then feeding that text into MiniMax to generate videos or images. This also represents MiniMax's status in many developers' minds: as a multimodal execution tool, M3 is usable, but not irreplaceable.
What truly dissatisfied users was M3's pricing scheme. On the day the model was released, MiniMax switched from its long-standing subscription-based Coding Plan to a new Token-based billing plan without notifying users in advance, and the official page's explanatory information was unclear. Many individual developers only discovered the rule change after logging in. Soon, people noticed that under identical usage intensity, quota consumption was far faster than expected. Dissatisfaction quickly fermented—some flocked to complaint platforms requesting refunds, some announced they would not renew, and others vented on social media. Shortly after, MiniMax issued an apology announcement, admitting that it had not fully communicated with users before the adjustment and that the handling of old users' weekly limits was inappropriate—"our work fell short"—and introduced a package of compensation measures.
But its market performance and reputation could not be fully restored. Several internal employees later described the months before M3's release to media: "All our attention was focused on the model's intelligence itself." Yan Junjie also admitted during the earnings call: "During M3's R&D process, we also had shortcomings in practice." The turning point came on July 31, when MiniMax released the open-source multimodal model H3. On Artificial Analysis's video model leaderboard, H3 ranked first globally in video editing capability, with a generation price of RMB 0.8/second (2K resolution), only one-third the price of comparable flagship video models. After H3's release, MiniMax's model capability reputation was somewhat reversed, as can be seen from the data. Yan Junjie stated at the earnings call: "In the three-plus weeks from H3's open-source release to August 26, H3 downloads exceeded 24 million, giving birth to over 300 public derivative models. It's one of the most downloaded models globally this year."
Meanwhile, the company is also undergoing continuous organizational adjustments. In late August, news emerged of the departure of A Dao (Miao Yuhang), head of MiniMax's Agent Engineering Department. He had been the public explainer of MiniMax's technical roadmap, responsible for M3.x, Agent, Audio, and Hailuo AI. His departure was also interpreted as a signal of MiniMax's shift from "engineer-driven" to "systematic organizational capability building." MiniMax is undergoing a brand-new transformation. Beyond maintaining technological leadership, a team of over 300 people suddenly becoming a hundred-billion-market-cap listed company means organizational capability, decision-making mechanisms, and external communication all still need time to integrate.
The Multimodal Gamble Market trust in the narrative of model capability still exists, but a more fundamental question is surfacing: which path to AGI should be taken? The large model industry is currently experiencing significant divergence in AGI routes. Zhipu's team has focused its main efforts on Coding capabilities, believing Agent is the key future direction; Moonshot AI's strategy is also to catch up with the world's frontier in pre-training levels through scaling, while vertically integrating model training with Agent products. In contrast, MiniMax insists on a full multimodal route. Yan Junjie clearly stated at the earnings call: "Unlike most companies, we have natively considered multimodal information when designing models, because we believe visual understanding and generation are important components of productivity. Multimodal generation is currently the second-largest market for AGI besides programming."
This judgment was not always so clear. In an earlier interview, Yan Junjie had said that while MiniMax's text-to-speech was the best in the industry, he felt the text model was most critical: "If the text model improves by 10 points, other modules naturally improve as well. The language model remains the most essential; everything else is naturally derived." In June this year, he further elaborated this logic at the developer conference: "Many peers are now mainly focusing on the Coding route, while we are one of the few companies insisting on investing simultaneously in Coding and content generation. We believe AGI's core value lies in improving society's productivity. Most of the work white-collar workers do on computers, besides information exchange, mainly involves two things: engineering creation centered on Coding, and creative expression centered on content generation. Therefore, content generation from voice to images to video is equally important to us."
In this regard, Yu Yang mentioned a judgment Yan Junjie had previously shared internally: the endgame of AI must be multimodal. To reach that endgame, exploration is necessary. If multimodal capabilities like video were cut now, by the time everyone approaches the endgame and understands how to do multimodal, it would be too late to catch up. "As a frontier company, you yourself are the leader; there's no one else to reference." This means MiniMax must simultaneously compete on multiple fronts—text, video, and audio—against opponents with more abundant resources.
Yan Junjie also emphasized at the earnings call that the company's most important growth source for the next phase remains model capability improvement, with plans to release new products like M3.1, M3 Pro, and H3.1. But the multimodal route has also raised practical concerns: will resources be dispersed? The prolonged absence of new models seems to confirm the market's worry. Yu Yang also admitted that multimodal business will have ups and downs during development. Looking at the current situation, MiniMax's multimodal business has been ongoing, just at different development paces and stages.
At the earnings call, Yan Junjie also responded to future compute allocation issues. The company will build a compute supply network through three paths: its own controllable core clusters, cloud vendor partnerships, and the Token Factory, balancing stability and flexibility. The text model is currently the highest priority for investment, with training resources approximately four times that of video models. Meanwhile, M3 and H3 are advancing domestic chip adaptation, with large-scale domestic compute clusters about to come online, gradually undertaking real production traffic to reduce unit Token costs.
But regardless of which path is chosen, AGI has always been Yan Junjie's goal. Huang Mingming, founding partner of MingShi Capital, an early MiniMax investor, also believes Yan Junjie genuinely has faith in AGI. In Yu Yang's view, for a large model company like MiniMax, the main trunk has always been model R&D; products like Xingye and Hailuo are more like branches growing from that trunk. All company development revolves around models. Using Xingye as an example: the core experience gap in such AI-native products always depends on the model—if the model is intelligent enough, the product experience is good. Rather than saying MiniMax is a product company, it has actually always been a model company.
On the wall of MiniMax's office building in Xuhui District, Shanghai, hangs the slogan "Intelligence with Everyone." This is MiniMax's vision since its founding and the goal Yan Junjie repeatedly mentions. Therefore, he believes a single model version doesn't represent everything. What matters more is whether the company can continuously define and iterate its own technical roadmap, whether it can continuously accelerate the improvement of intelligence density, and whether it can deliver higher levels of intelligence to more users at lower unit costs. But before the endgame Yan Junjie envisions arrives, when will MiniMax's next sufficiently powerful new model come? The market is still waiting.
Comments