How MiniMax Turned Explosive AI Usage Into an Enterprise-Led Business

MiniMax’s business changed direction sharply between 2025, February 2026, and August 2026. It began as a company with substantial consumer-facing AI products and model development ambitions; by August, its reported commercial momentum was increasingly tied to enterprise and developer inference usage.

The headline numbers capture the acceleration. MiniMax reported $116.6 million in revenue for the first half of 2026, up 283.1% year over year and already exceeding its full-year 2025 revenue of $79.0 million. 61819 Management also said its annualized recurring revenue (ARR) had surpassed $800 million in August, compared with more than $150 million in February. 51011

That does not mean MiniMax had already recognized $800 million in accounting revenue. ARR extrapolates a current recurring-revenue pace over a year, so it is better understood as a signal of momentum than as a substitute for reported sales. 10

The business moved from consumer-led to enterprise-led

The clearest change was the revenue mix. Open-platform and other AI-based enterprise services generated $73.9 million in the first half, a 703.1% year-on-year increase. Their share of revenue rose to 63.4%, compared with 30.3% in the first half of 2025. 6713

Management said B2B services accounted for more than 80% of August ARR. 510 In practical terms, MiniMax was no longer relying primarily on consumer subscriptions or standalone AI applications. Its platform, API, and inference business had become the largest source of incremental commercial growth.

The quarterly trend points in the same direction. Second-quarter revenue increased 81.8% from the first quarter, while token consumption in July reached about 20 times January’s level. 51533 Those indicators suggest that usage was expanding inside applications and workflows, not merely through one-off experimentation.

Consumer products continued to grow

The shift toward B2B did not mean MiniMax abandoned the consumer market. Revenue from its AI-native products increased 100.9% year over year to $42.6 million in the first half of 2026. 7

Products such as Talkie and Hailuo AI remained part of the company’s growth story, providing distribution, user activity, and potential demand for new multimodal experiences. But the financial mix shows that consumer growth was no longer the dominant commercialization narrative. Enterprise and developer services generated more revenue and were contributing an even larger share of the recurring run rate by August.

Why inference demand accelerated

Several developments helped connect MiniMax’s models to more sustained commercial usage.

M3 targeted coding and agentic work

MiniMax released M3 with an emphasis on coding, agentic reasoning, tool use, multimodal inputs, and long-context work. The company’s own release materials describe frontier-level results on specialized coding and agent benchmarks, including SWE-Bench Pro and Terminal-Bench 2.1. 1722

Those capabilities matter commercially because production agents perform multi-step tasks: they may read files, call tools, write code, test outputs, and revise their work. Such workflows can generate much more inference demand than a simple conversational exchange. The July token-consumption figure—20 times January’s level—fits MiniMax’s argument that agent-oriented workloads are increasing the amount of model usage per task. 515

Open models expanded distribution

MiniMax also released H3, which it describes as a general-purpose multimodal video model that works across text, images, video, and audio. The model is positioned for commercial content creation and is available as an open model. 1720

An open distribution strategy can put models in front of developers who may later build products, pay for hosted inference, or bring larger workloads to the platform. The provided evidence supports H3’s role as a developer-distribution and ecosystem initiative, but does not by itself establish how much revenue H3 generated.

Adoption broadened beyond individual users

MiniMax says its models and AI-native products have served more than 300 million individual users and more than two million enterprises and developers across global markets. 2127

These are company-reported cumulative reach figures rather than a disclosed count of paying customers. Still, they illustrate the scale of the funnel MiniMax is trying to convert into recurring platform usage: consumer exposure creates distribution, while enterprise and developer deployments create higher-value inference demand.

Economics improved, but profitability remained distant

MiniMax’s gross economics improved during the first half. Gross profit rose 464.8% year over year to $20.8 million, and gross margin increased from 12.1% to 17.9%. 719

The company attributed the improvement partly to infrastructure efficiency and lower-cost deployment. Management also said its self-built infrastructure had reached a 97% Effective Training Ratio, although that operational metric should not be confused with gross margin or profitability. 5

Operating investment remained heavy. Research and development expense increased 138.8% year over year to approximately $296.9 million, reflecting continued spending on foundation models and multimodal capabilities. 4750 MiniMax reported a first-half loss of $358.0 million, narrower than the $402.2 million loss recorded a year earlier. 1419

The result is a mixed commercialization picture: revenue and gross margin improved quickly, but the company was still far from profitable. A central question is whether infrastructure gains and greater paid usage can continue to outpace the cost of training, inference, and expansion.

The commercialization flywheel MiniMax is proposing

MiniMax’s strategy rests on a positive feedback loop rather than a simple trade-off between model intelligence and low cost:

1. More capable models complete more valuable tasks, especially coding and agent workflows.
2. Better task performance encourages companies and developers to move from testing to production deployment.
3. Production deployment increases token consumption and recurring inference revenue.
4. Higher volume and more efficient infrastructure can lower the cost of serving each token.
5. Lower prices make it economical to use models in more applications, generating additional volume and revenue.
6. That revenue can support further model training and infrastructure investment, improving the next generation of products.

The early evidence supports parts of this loop: M3 was positioned around coding and agentic work, token consumption rose sharply, enterprise revenue grew 703.1%, and gross margin improved. 67151722 But the complete flywheel remains management’s thesis, not a proven long-term outcome. Sustaining low prices while funding heavy research and narrowing losses will be the harder test.

What changed between 2025 and August 2026?

The comparison can be summarized in three stages:

• 2025: MiniMax generated approximately $79.0 million in full-year revenue, with enterprise-related revenue representing a much smaller share of the business than in 2026. 618
• February 2026: ARR surpassed $150 million, signaling that model usage was beginning to translate into a larger recurring platform business. 4244
• August 2026: ARR exceeded $800 million; first-half revenue had already surpassed full-year 2025 revenue; enterprise services represented 63.4% of first-half revenue and more than 80% of August ARR. 571018

The strongest conclusion is not simply that MiniMax grew faster. It commercialized differently. Its center of gravity moved toward APIs, developer adoption, enterprise deployments, and recurring inference consumption, while consumer products remained an important but comparatively smaller growth channel. The company had built a more credible enterprise platform story by August—but not yet a profitable one.