Introduction: The "Chinese OpenAI" is now pivoting to the B2B market.
Recently, MiniMax open-sourced version v0.4.12 of its programming tool, MiniMax Code CLI, under the MIT license for global developers. This command-line tool achieved a 76.7% task pass rate in the Frontier Harness Eval benchmark, can be installed with a single command, and is free for anyone to use, modify, and distribute.
No user stories, no product launch events, and no narrative packaging of a "C-end hit." The purpose of the open-source release is stated straightforwardly: to allow developers to examine the tool's invocation permissions and build reliable enterprise-grade applications.
This is a move targeting the B2B market.
But if we turn the clock back seven months, MiniMax's narrative was entirely different. On January 9, 2026, the company's shares surged 109% on the first day of its listing on the Hong Kong Stock Exchange, with a 1,837x oversubscription, pushing its market capitalization past HKD 100 billion. The reason the capital market favored it at the time was a global C-end story: Talkie had accumulated 300 million users overseas, Xingye was doing AI companionship in China, and Hailuo AI was focused on video generation—a "Chinese OpenAI" growing from the application layer upwards.
Seven months later, this story is being quietly replaced.
The semi-annual report shows that in the first half of 2026, MiniMax's B2B revenue grew 703% year-on-year, with its share of total revenue surging from 30.3% to 63.4%. Management disclosed at the earnings call that the August ARR (Annual Recurring Revenue) exceeded USD 800 million, of which the B2B segment contributed about 80%. The number of enterprise clients and developers broke through 2 million, 10 times the figure at the end of last year.
A company that started with emotional companion applications has turned into a "Token-selling platform" in less than a year.
The C-end Story Has Reached Its Limit
MiniMax's C-end story was once one of the most attractive narratives in China's AI industry.
In June 2023, Talkie was launched, tapping into AI role-playing. Users can create their own characters, adjust their appearance and personality, and interact with them via text or voice. For users who use it continuously for a week, the payment rate and retention rate are around 80%. The product team of nearly 300 people put a lot of effort into UI details: the gold and black color scheme pursues a sense of luxury, the dialogue box automatically sets a bracket style to facilitate users in inputting voice-overs, and star elements appear just right in the characters' eyes and payment nodes.
This product has created a growth myth globally, consistently ranking among the top three in downloads for companion AI applications, alongside US unicorns Character.AI and Replika. The prospectus shows that in the first three quarters of 2025, AI-native products contributed 71.1% of revenue, and the number of paying users grew 15 times in two years.
But turning to the segment data in the prospectus, the other side is exposed: the gross profit margin of C-end AI-native products is only 4.7%.
For every USD 1 of revenue generated, the gross profit is less than 5 cents. After deducting product operations, platform commissions, and inference costs, this business makes almost no money.
Meanwhile, the growth curve is flattening. In the fourth quarter of 2025, the C-end revenue growth rate was halved from nearly 150% to 82%. According to data from the AI Product Ranking, as of August 2026, Talkie's global MAU (Monthly Active Users) was 31.07 million, a month-on-month decrease of 5.93%; the MAU of the domestic version, Xingye, was 3.38 million, a month-on-month decrease of 5.73%. The paid conversion rate is only about 6.9%, and the ARPPU (Average Revenue Per Paying User) is just USD 5.
Out of over 200 million cumulative users, very few are truly willing to keep paying.
This is not just a problem for MiniMax. The entire AI companionship track faces the same structural dilemma: users are willing to chat with AI characters, but they are not willing to pay too much for it.
The C-end story has hit the ceiling. It is not that the product is bad, but that the economic model of this business itself has fundamental flaws.
Pivoting to the Token Business
MiniMax's pivot came quickly.
In the first half of 2026, B2B open platform revenue surged from USD 9.2 million to USD 73.9 million, a year-on-year growth of 703%, making it the largest source of revenue overnight. By August, the B2B segment's share in ARR had reached about 80%.
The underlying logic of this transformation is that the way models are used has changed. In the past, when a user chatted with AI, one question corresponded to a few rounds of dialogue, with limited Token consumption. After Agents became popular, users delegate tasks to AI, and the AI decomposes, searches, plans, executes, and verifies on its own. A single task can generate dozens of times the Token consumption.
Data shows that the domestic daily average Token invocation volume was about 100 billion in early 2024, reached 100 trillion by the end of 2025, and broke through 140 trillion in March 2026, growing by over a thousand times in two years.
The explosive growth of MiniMax's B2B business is essentially riding the fast train of Agents. The open-sourced Code CLI is an extension of this strategy—it is not a product for ordinary consumers, but a programming tool for developers, aiming to bring more developers into the MiniMax ecosystem, embed model capabilities into more enterprise workflows, and then monetize through continuous Token consumption.
From "selling characters" to "selling tools," from "selling experiences" to "selling computing power," MiniMax is completely changing its face.
But "Selling Shovels" Is Not as Profitable as Imagined
However, the B2B market is not a safe haven.
In the first half of 2026, MiniMax's overall gross profit margin was 17.9%. Although this improved from 12.1% in the same period last year, it was far below the 25.4% for the full year of 2025. Revenue grew by 283%, and the cost of revenue grew by 258%. The gross profit margin did not expand significantly due to the increased proportion of B2B business.
The direct cause of the gross profit margin decline was a failed price hike. On the day the M3 model was released on June 1, MiniMax changed its billing model from per-request to per-Token without any prior warning, and the Starter package jumped from CNY 29 to CNY 49. Communities like V2EX were immediately flooded with complaints, and users concentrated their feedback on the Black Cat Complaint platform. The stock price plunged 15.71% that day. Facing the backlash from developers and the price war pressure from competitors like DeepSeek, MiniMax was forced to announce a permanent 50% cut in API prices on June 15.
This exposes a deep-seated problem: most of MiniMax's B2B clients are small and medium-sized developers who are price-sensitive. They choose MiniMax largely because of its "high cost-performance ratio" rather than its "strongest capabilities." Once competitors lower their prices, the customer migration cost is extremely low.
In contrast, Zhipu AI, backed by its model capabilities, raised the pricing of its CodingPlan code package by 83% a quarter earlier, while still achieving volume growth. Zhipu AI's GLM-5.3 scored 60 out of 63 in the AA Comprehensive Intelligence Index, tying with ClaudeFable5 and GPT-5.6 at the global frontier; whereas MiniMax's M3 scored only 45, ranking outside the top ten. According to an analysis report by iAnalysis, among the first tier of Token revenue, the "CNY 10 Billion Club", ByteDance, Alibaba, and Zhipu AI were shortlisted, while MiniMax is in the second tier (around CNY 5 billion), alongside DeepSeek, Kimi, and Kling.
More notably, Zhipu AI's API gross profit margin has turned positive from about -0.4% to 24.6%, the inference cost per Token has dropped by about 80% compared to the beginning of the year, and the API revenue corresponding to every CNY 1 invested in computing power has increased by about 14 times year-on-year.
Another detail worth noting is the statistical caliber of ARR. The "USD 800 million ARR" disclosed by MiniMax's management is calculated by annualizing a certain highest weekly data point in August, rather than annualizing monthly data. If calculated using the conventional monthly caliber, the actual figure would be lower than this statement.
Of course, MiniMax is not without highlights. The H3 video generation model open-sourced at the end of July saw over 24 million downloads in three weeks, spawning more than 300 public derivative models. H3Max takes less than 3 seconds to generate a 5-second video, achieving for the first time a video generation speed faster than the playback speed.
However, the breakthrough in video models solves the problem of dissemination; only the quality and stability of the production pipeline can solve the retention of paying users. Moreover, H3 focuses on cost-effectiveness with relatively low prices, and its gross profit margin is lower than that of ByteDance's Seedance and Kuaishou's Kling, which are also video models.
MiniMax is currently in an awkward middle ground: the C-end's over 200 million users failed to support a high-gross-margin model, and the B2B's 2 million clients have yet to prove their loyalty; its text model capabilities lag behind Zhipu AI and Kimi, and although its video model is popular in the open-source community, it yields meager profits. The full-modality in-house development route means resources are diluted.
From the C-end story to the B2B narrative, MiniMax is indeed heading towards a different story. But the core suspense of this story remains unchanged: whether "selling characters" or "selling Tokens," the company has yet to prove that it can establish true pricing power.
And pricing power is the ultimate yardstick that determines how much an AI company is ultimately worth.