Introduction: The entity dubbed the "Chinese counterpart of OpenAI" is now making strides into the B-end market.

Recently, MiniMax released version v0.4.12 of its programming tool, MiniMax Code CLI, as open-source software under the MIT license, catering to developers worldwide. This command-line tool achieved a 76.7% task pass rate in the Frontier Harness Eval evaluation and can be installed with a single command, granting users the freedom to utilize, modify, and distribute it at no cost.

There are no user anecdotes, no product debuts, and no "C-end sensations" packaged in a narrative. The rationale behind open-sourcing is straightforward: to empower developers to scrutinize the tool's access permissions and construct dependable enterprise-level applications.

This move is clearly targeted at the B-end market.

However, if we rewind seven months, MiniMax's narrative was vastly different. On January 9, 2026, the company witnessed a remarkable 109% surge on its debut at the Hong Kong Stock Exchange, with an oversubscription rate of 1,837 times and a market capitalization surpassing HK$100 billion. Back then, the capital markets were enthralled by its potential for a global C-end story: Talkie had amassed 300 million users overseas, Xingye provided AI companionship in China, and Hailuo AI handled video generation—a "Chinese version of OpenAI" evolving from the application layer upwards.

Seven months have elapsed, and this narrative is quietly being supplanted.

The interim report reveals that MiniMax's B-end revenue skyrocketed by 703% year-over-year in the first half of 2026, with its share of total revenue soaring from 30.3% to 63.4%. During the earnings call, management disclosed that August's Annual Recurring Revenue (ARR) exceeded $800 million, with the B-end contributing approximately 80%. The number of enterprise clients and developers surpassed 2 million, a tenfold increase from the end of the previous year.

A company that initially focused on emotional companionship apps has transformed into a "token-selling platform" in less than a year.

01 The C-end Story Has Reached Its Limit

MiniMax's C-end narrative was once among the most captivating in China's AI industry.

In June 2023, Talkie was launched, focusing on AI role-playing. Users could create their own characters, customize appearances and personalities, and interact via text or voice. Users who engaged continuously for a week exhibited around 80% payment and retention rates. The product team, comprising nearly 300 individuals, invested significant effort into UI details: a gold-and-black color scheme for a luxurious feel, dialogue boxes automatically set in brackets for easy aside input, and star elements strategically placed in character eyes and payment nodes.

This product achieved global acclaim, consistently ranking among the top three in downloads for companionship AI apps, alongside U.S. unicorns Character.AI and Replika. The prospectus indicates that AI-native products contributed 71.1% of revenue in the first three quarters of 2025, with paying users growing 15-fold in two years.

However, a closer look at the prospectus's segmented data reveals another perspective: the gross margin for C-end AI-native products was a mere 4.7%.

For every $1 in revenue, the gross profit was less than 5 cents. After deducting product operations, platform fees, and inference costs, the business barely turned a profit.

Meanwhile, the growth curve was flattening. In the fourth quarter of 2025, C-end revenue growth plummeted from nearly 150% to 82%. According to AI Product Rankings, as of August 2026, Talkie's global Monthly Active Users (MAU) were 31.07 million, down 5.93% month-over-month; Xingye's domestic MAU was 3.38 million, down 5.73% month-over-month. The payment conversion rate was only about 6.9%, with an Average Revenue Per Paying User (ARPPU) of just $5.

Among over 200 million cumulative users, truly willing and sustained payers were scarce.

This issue is not unique to MiniMax. The entire AI companionship sector faces the same structural dilemma: users are willing to chat with AI characters but reluctant to pay much for it.

The C-end story has hit a ceiling. Not because the product is inferior, but because the business's economic model is inherently flawed.

02 Shifting to Token Business

MiniMax's pivot was swift.

In the first half of 2026, B-end open platform revenue soared from $9.2 million to $73.9 million, up 703% year-over-year, becoming the primary revenue source. By August, the B-end accounted for about 80% of ARR.

The underlying logic of this transition is a shift in how models are utilized. Previously, a user chatting with AI would generate a few rounds of dialogue per question, resulting in limited token consumption. After agents gained popularity, users delegated tasks to AI, which then broke them down, searched, planned, executed, and validated them independently, potentially generating dozens of times more token consumption per task.

Data indicates that China's average daily token calls were around 100 billion in early 2024, reaching 100 trillion by the end of 2025, and surpassing 140 trillion in March 2026, growing over a thousandfold in two years.

MiniMax's B-end explosion essentially capitalized on this agent wave. The open-sourced Code CLI is an extension of this strategy—not a product for ordinary consumers but a programming tool for developers, aiming to onboard more developers into the MiniMax ecosystem, embed model capabilities into more enterprise workflows, and then monetize through sustained token consumption.

From "selling characters" to "selling tools," from "selling experiences" to "selling computing power," MiniMax is undergoing a complete transformation.

03 But "Selling Shovels" Isn't as Profitable as Imagined

Yet the B-end is no safe haven.

In the first half of 2026, MiniMax's overall gross margin was 17.9%, up from 12.1% year-over-year but far below 25.4% for all of 2025. Revenue grew 283%, while the cost of revenue grew 258%, leaving little room for gross profit expansion despite the rising B-end share.

The direct cause of the margin decline was a failed price hike. On June 1, the day the M3 model launched, MiniMax abruptly switched its billing model from per-request to per-token, with the Starter package jumping from 29 yuan to 49 yuan. Complaints flooded communities like V2EX, with users concentrating feedback on Black Cat Complaints. The stock plummeted 15.71% that day. Facing developer backlash and price war pressure from rivals like DeepSeek, MiniMax was forced to halve API prices permanently on June 15.

This exposed a deeper issue: MiniMax's B-end clients are mostly price-sensitive small-to-medium developers who chose MiniMax largely for "cost-effectiveness" rather than "superior capability." Once competitors cut prices, client migration costs are extremely low.

In contrast, Zhipu, bolstered by model capabilities, raised its CodingPlan code package pricing by 83% a quarter earlier while still experiencing volume growth. Zhipu's GLM-5.3 scored 60 (out of 63) in the AA Composite Intelligence Index, tying with ClaudeFable5 and GPT-5.6 at the global frontier; MiniMax's M3 scored only 45, ranking outside the top ten. According to iResearch, among the first-tier "$10 billion club" in token revenue, ByteDance, Alibaba, and Zhipu made the cut, while MiniMax was in the second tier (5 billion yuan range), alongside DeepSeek, Kimi, and Kling.

More notably, Zhipu's API gross margin turned positive from about -0.4% to 24.6%, with unit token inference costs dropping about 80% year-over-year and API revenue per yuan of computing power invested rising about 14-fold year-over-year.

Another noteworthy detail is the ARR calculation method. MiniMax management's "$800 million ARR" was annualized based on a peak weekly figure in August, not monthly data. Using the standard monthly method, the actual figure would be lower.

Of course, MiniMax isn't without its bright spots. The H3 video generation model open-sourced in late July surpassed 24 million downloads in three weeks, spawning over 300 public models. H3Max generates a 5-second video in under 3 seconds, achieving video generation speed exceeding playback speed for the first time.

However, the video model's breakthrough addresses dissemination issues; only production pipeline quality and stability can solve paying user retention. Moreover, H3 focuses on cost-effectiveness, with lower pricing and gross margins than ByteDance's Seedance and Kuaishou's Kling.

MiniMax is stuck in an awkward middle ground: its 200 million C-end users failed to sustain high-margin models, while its 2 million B-end clients haven't proven loyalty; its text model lags behind Zhipu and Kimi, and its video model, though open-source popular, is low-margin. The all-modal self-research route means resources are spread thin.

From C-end stories to B-end narratives, MiniMax is indeed heading toward another story. But the core question remains unchanged: whether "selling characters" or "selling tokens," the company has yet to prove it can establish genuine pricing power.

And pricing power is the ultimate yardstick determining how much an AI company is truly worth.