MiniMax market cap shrinks ~$42B from peak as multi-modal AI strategy faces investor skepticism

Source: China Entrepreneur Magazine Reporter: Kong Yuexin Editor: Ma Jiying Cover Image Source: Visual China

Market Reaction: Why Isn't the Secondary Market Buying In?

Since September, MiniMax has been making strategic moves in overseas markets.

On September 16, under an AI learning initiative launched by Singapore's SkillsFuture Singapore agency, three MiniMax products were selected as the only Chinese domestic large-model representatives, 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, built on MiniMax's M3 base model and pre-trained with over 1 trillion tokens of native Arabic data. It achieved the highest average score across seven Arabic benchmark tests.

Driven by these developments, MiniMax's stock price rose for two consecutive days on September 17 and 18, with a nearly 19% surge at the Hong Kong market close on the 18th. As of the close on September 21, MiniMax shares stood at HKD 298.

Compared to its peak of HKD 1,330 per share, the market is now more focused on 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 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 M3 Pro is expected to have approximately 3 trillion parameters, equipped with the MSA 2.0 architecture, and computing efficiency is expected to improve by about three times."

However, as of now, there has been no news of M3.1's launch.

Meanwhile, the iteration pace of domestic large models has accelerated noticeably since the second half of 2026:

• Zhipu released GLM-5.3 on August 14
• Kimi launched K3 with 2.8 trillion parameters on July 17
• DeepSeek V4 Pro officially launched on August 13

Although MiniMax released its multimodal video model H3 on July 31, text model capabilities remain the primary focus under current evaluation frameworks. M3, released hastily and with insufficient preparation, found itself sandwiched between GLM-5 and K3—both of which excelled in areas like coding and agent capabilities. Combined with strong developer backlash over pricing, M3's position in the competitive landscape became particularly awkward.

From a product iteration logic perspective, releasing a superior model to cover the previous misstep would be the optimal solution. However, the company's delay in launching M3.1 and M3 Pro has raised external doubts about MiniMax's model capabilities.

Currently, when investors think of Zhipu, they think of coding; Kimi is associated with agent clusters; DeepSeek is known for text focus and pricing advantages. These labels and partial leadership in model capabilities have given these companies clear valuation anchors in the capital market and clearer user recognition. But MiniMax's multimodal strategy has not yet formed an absolute leading advantage in any single capability or pricing. The market perceives its positioning as consistently vague, causing its valuation framework to waver.

In June 2026, JPMorgan downgraded MiniMax from "Overweight" to "Neutral" and significantly lowered its target price. The core reason: MiniMax's flagship M3 model lacks pricing power, and its model capabilities are still in a catch-up phase. This means MiniMax must deliver a sufficiently powerful model to prove its strategic direction is correct.

In reality, the questions MiniMax needs to answer extend far beyond the model itself.

From B2C to B2B

On August 26, MiniMax released its first interim results since listing.

First Half 2026 Financial Highlights:

• Total revenue: $117 million, up 283.1% year-over-year
• Half-year revenue already 1.5 times full-year 2025 revenue
• Q2 revenue: up 81.8% quarter-over-quarter
• July token consumption: 20 times January levels
• August ARR (Annual Recurring Revenue): exceeded $800 million
• Gross profit: $20.81 million, up 464.8% year-over-year
• Gross margin: improved from 12.1% to 17.9%, up 5.8 percentage points

However, other figures tell a different story:

• R&D expenses in H1: $297 million, up 138.8% year-over-year
• Net loss for the period: $358 million, narrowing 11% from $402.2 million in the same period last year
• Adjusted net loss (excluding share-based payments, fair value changes in financial liabilities, and listing expenses): $293 million, expanding 111.2% year-over-year
• Adjusted net loss rate: narrowed from approximately 455.9% to approximately 251.4%—an apparent improvement, but primarily due to a larger revenue base; the absolute loss is still accelerating

More noteworthy than profit/loss figures is the dramatic shift in revenue structure.

In the first half of 2025, MiniMax's AI-native products (Hailuo AI, Talkie/Xingye, etc.) accounted for approximately 70% of revenue, serving as the company's absolute pillar. By the first half of 2026, B2B open platform and enterprise service revenue surged from $9.2 million to $73.9 million, up 703.1% year-over-year, jumping from 30.3% to 63.4% of total revenue, becoming the largest revenue source. AI-native product revenue, while growing 100.9% year-over-year to $42.6 million, fell to 36.6% of total revenue. By August, B2B accounted for approximately 80% of ARR, with B2C at only about 20%.

Photo: Kong Yuexin

Within six months, a company known to ordinary users for its consumer products had completely inverted its revenue structure.

However, this shift was not without warning. At least on the coding front, internal discussions at MiniMax began much earlier than the outside world expected. On October 27, 2025, MiniMax officially open-sourced and launched MiniMax M2, touted as purpose-built for agents and code.

Yu Yang, an entrepreneur and former MiniMax employee, told China Entrepreneur: "The company's judgment on coding actually emerged as early as 2025 when Claude first appeared. At that time, IO (Yan Junjie) had already shared his views at the CD meeting." The CD meeting is MiniMax's weekly internal meeting held every Friday at noon, where Yan Junjie reports on business developments, answers employee questions, and shares his views on AI frontier developments.

Therefore, in Yu Yang's view, MiniMax's commercialization push toward B2B was foreseeable early on. "Because entering the coding赛道 means serving B2B clients."

At a MiniMax developer offline 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 1 to 2 million people in China could write code—seemingly not a broad enough market. But that has clearly changed. AI coding can indeed give more ordinary people productivity.

However, early布局 does not automatically translate into market advantage.

MiniMax's model parameters and capabilities in the coding赛道 have not formed a sufficiently strong cognitive barrier among developers.

This cognitive lag quickly reflected in the capital market. When market perception of model capabilities failed to match technological investment, the valuation logic began to waver. The secondary market faced lock-up expiration pressure and valuation regression. Although over 80% of pre-IPO and cornerstone shareholders publicly stated they would continue holding, and two major strategic shareholders—Alibaba and miHoYo—explicitly expressed support, the company's stock price still experienced剧烈 volatility, falling over 80% from its peak.

After MiniMax's August 26 earnings release, the stock price rose only 3.83% the following day. On August 31, following the launch of a livestream built on H3 Max, MiniMax's stock price rapidly climbed, rising nearly 20% intraday and closing up 16.18%. It continued its upward trend on September 1, accumulating over 20% gains over two trading days.

This indicates that the market still believes in the narrative of model capability. Yan Junjie reiterated this point during the earnings call. 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 in the second half of last year, to more autonomous and creative completion of long-term tasks, to the future expectation of end-to-end delivery of reliable results, we continue to pursue higher intelligence."

M3's "Derailment" and H3's "Turnaround"

On June 1, MiniMax released M3. The model has a total parameter count of 428B, with 23B activated parameters, employing a Mixture of Experts (MoE) architecture, and natively supports million-level token context.

Within the competitive landscape at the time, this parameter scale appeared somewhat尴尬.

When M3 was released, several trillion-parameter models had already been deployed domestically, including Qwen3-Max-Thinking, DeepSeek V4 Pro, and Kimi K2.6. Overseas, GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.7 were competing simultaneously. Trillion-level parameters were becoming the entry ticket for top players. Two months later, Kimi K3 open-sourced with 2.8 trillion parameters, and Qwen-3.8 Max reached 2.4 trillion, further widening the gap.

Source: Visual China

Before its release, M3 carried high expectations, and internally, the company hoped the model's capabilities would reach a new level. However, after launch, many developers reported that M3's actual performance fell short of expectations. An insider admitted that M3's launch was somewhat "hasty" or "rushed."

Alex, an individual developer, said his primary daily-use model is Kimi K3. When his quota runs out, he tries Zhipu's GLM, with MiniMax as a backup option.

"MiniMax gives me the feeling that its capabilities aren't that strong, but it works steadily." Alex explained his selection logic: He is not a software engineer and lacks strong discernment. When asking AI to perform functions, he often cannot understand most of the questions the model asks him. In such cases, he prefers to delegate decision-making to the AI, letting it decide how to proceed, and then verify the results. "So I need a stronger model, not a dumb but fast one, because the latter would significantly increase the frequency of my homework checks." Therefore, he prioritizes Kimi or stronger SOTA models.

Chen Lijun, founder of Yita Industrial Intelligent Technology, requires higher model capability and code accuracy for his work providing software to the industrial sector. M3's performance could not meet his delivery requirements. He believes the root issue lies in parameter count. Using Qwen's models as an example: "3.7 Max and Plus have the same total parameter count, but Plus is multimodal. Adding multimodal parameters at the same parameter level causes a sharp decline in coding quality. Switching to the trillion-parameter 3.8 Max, although multimodal, its capabilities are on par with or even higher than 3.7." For Chen Lijun, models specifically trained for coding, or pure text models within the same parameter class, are generally stronger than multimodal models.

For Chen Lijun, M3's problem is insufficient parameters.

He calculated: "What really consumes tokens is rework. I'd rather spend more tokens on upfront research, get each module right the first time, and then call them later—that yields the highest returns. If you make something and then ask AI to fix it, it might delete previous work while making changes. It's very uncontrollable—I've personally fallen into that deep pit." In his view, a model's "cost-effectiveness" doesn't hold up in daily workflows: money saved on low prices may be doubled in rework costs.

However, M3 is not necessarily unusable. Developer Xiao Feng emphasized that methodology matters more than the model itself. In his view, everyone uses models differently. Some people may not even use built-in skills, and the problem may not lie with the model. "You give the AI the big direction, and it helps you implement it. But before execution, you need to plan first—generate a to-do list, check whether the content meets requirements, and then have it execute based on the list." He also mentioned that deleting 80% of Claude Code's prompts can still achieve results; the key lies in whether the requirement description is clear and structured. "Choose the direction first, then execute. You can't rely entirely on AI."

Xiao Feng said he uses MiniMax daily for coding, image, and image recognition tasks—for example, generating text with DeepSeek and then feeding it into MiniMax to generate videos or images.

This reflects 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 of the model's release, MiniMax switched its long-standing subscription-based Coding Plan to a new token-based billing plan. However, this change was not communicated to users in advance, and the official page's explanations were unclear. Many individual developers only discovered the rule change after logging in. Soon, users found that at the same usage intensity, token consumption far exceeded expectations. Dissatisfaction quickly escalated: some flocked to complaint platforms demanding refunds, others announced they would not renew subscriptions, and vented on social media.

MiniMax quickly issued an apology, acknowledging that it had not adequately communicated with users before the adjustment and that the handling of old users' weekly limits was inappropriate—"our work was insufficient"—and rolled out a compensation package. However, its market performance and reputation could not be fully restored.

Multiple employees later described the months before M3's launch to the media: "Our entire focus was on the model's intelligence itself." Yan Junjie also admitted during the earnings call: "During M3's R&D process, we also had practical shortcomings."

The turning point came on July 31, when MiniMax released the open-source multimodal model H3. On the Artificial Analysis 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 of comparable flagship video models.

After H3's release, MiniMax's model capability reputation was somewhat reversed, as reflected in the data. Yan Junjie stated during the earnings call: "From H3's open-source release to August 26, over three weeks, H3 had been downloaded more than 24 million times, with over 300 publicly available derivative models. It is one of the most downloaded models globally this year."

Meanwhile, organizational adjustments were also underway.

In late August, MiniMax's Agent Engineering Department head, Adao (Miao Yuhang), reportedly left the company. He had been a public explainer of MiniMax's technical路线, responsible for M3.x, Agent, Audio, and Hailuo AI. His departure has been interpreted as a signal of MiniMax's transition from "engineer-driven" to "systematic organizational capability building."

MiniMax is undergoing a comprehensive transformation. Beyond maintaining technological leadership, a team of over 300 people has suddenly become a listed company with a market cap exceeding HKD 100 billion. Organizational capabilities, decision-making mechanisms, and external communication all require time to磨合.

The Multimodal Bet

Market trust in the narrative of model capability still exists, but a more fundamental question is emerging: Which path leads to AGI?

The large-model industry is currently experiencing significant divergence in AGI routes. Zhipu's team focuses primarily on coding capabilities, believing agents are the key future direction. Moonshot AI's strategy involves catching up with global frontiers in pre-training through scaling, while vertically integrating model training with agent products.

In contrast, MiniMax insists on a full multimodal route. Yan Junjie clearly stated during the earnings call: "Unlike most companies, we have always 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 after programming."

MiniMax founder and CEO Yan Junjie | Source: China Entrepreneur Photo Library

This judgment was not always so clear.

In an earlier interview with China Entrepreneur, Yan Junjie said that while MiniMax's text-to-speech was the best in the industry, he believed text models were most critical. "If the text model improves by 10 points, other modules will naturally improve too. Language models are still the most fundamental; everything else is a natural derivative."

In June this year, he further elaborated this logic at a developer conference: "Many peers are currently focusing on the coding route, while we are one of the few companies investing simultaneously in both coding and content generation. We believe the core value of AGI lies in improving social productivity. Most of the work white-collar workers do on computers, aside from information exchange, mainly involves two things: engineering creation centered on coding, and creative expression centered on content generation. Therefore, content generation—from voice and images to video—is equally important to us."

Yu Yang mentioned a judgment Yan Junjie had shared internally: The endgame of AI must be multimodal. To reach that endgame, exploration is necessary. If multimodal capabilities like video are cut now, by the time the endgame approaches and everyone understands how to do multimodal, it will be too late to catch up. "As a frontier company, you are yourself a leader; there is no one else to reference."

This means MiniMax must simultaneously compete on multiple fronts—text, video, audio—against rivals with more abundant resources.

Yan Junjie also emphasized during the earnings call that the company's most important growth driver in the next phase remains model capability improvement, with new products like M3.1, M3 Pro, and H3.1 to follow.

However, the multimodal route also raises a practical concern: Are resources being spread too thin? The delayed launch of new models seems to confirm this market worry.

Yu Yang acknowledged that multimodal business will experience ups and downs during development. From the current situation, MiniMax's multimodal business has been ongoing, but each area is at a different stage and pace of development.

During the earnings call, Yan Junjie also addressed future computing power allocation. The company will build its computing power supply network through three paths: self-controlled core clusters, cloud vendor partnerships, and Token Factory, balancing stability and flexibility. Text models are 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 computing power clusters soon to come online, gradually taking on real production traffic to reduce unit token costs.

But regardless of which path is chosen, AGI remains Yan Junjie's goal.

Huang Mingming, founding partner of Mingshi Venture Capital and an early MiniMax investor, believes Yan Junjie genuinely has faith in AGI.

In Yu Yang's view, for a large-model company like MiniMax, the主干 remains model R&D. Products like Xingye and Hailuo are more like branches growing from the主干. All company development revolves around the model. Using Xingye as an example: The core experience gap for 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 Shanghai's Xuhui District hangs the slogan: "Intelligence with Everyone." This has been MiniMax's vision since its founding and a goal Yan Junjie repeatedly mentions. Therefore, he believes that a single model version does not represent everything. What matters more is whether the company can continuously define and iterate its technical路线, whether it can accelerate the improvement of intelligence density, and whether it can deliver higher-level intelligence to more users at lower unit costs.

But before Yan Junjie's envisioned endgame arrives, the market is still waiting: When will MiniMax's next sufficiently powerful new model come?

News Hotline & Submission Email: tougao@iceo.com.cn

Disclaimer: The above content represents only the author's views or positions and does not represent the views or positions of Sina Finance & Economics Headlines. For any issues related to content, copyright, or other matters requiring contact with Sina Finance & Economics Headlines, please do so within 30 days of the above content's publication.