Chinese AI unicorn MiniMax reported its first annual results since going public, delivering growth that exceeded expectations. Full-year 2025 revenue reached $79.04 million, a 159% year-over-year increase, approximately 10.6% higher than the Bloomberg consensus estimate of $71.4 million. Annual Recurring Revenue (ARR) surpassed $150 million by February 2026, indicating a significant acceleration in the company's commercialization pace.

Profitability improved concurrently. Full-year gross profit surged 437% to $20 million, with the gross margin expanding from 12.2% in 2024 to 25.4%. The full-year adjusted net loss was $250 million, remaining largely flat compared to the previous year.

During the earnings call, Founder and CEO Yan Junjie outlined the company's strategic evolution: MiniMax is transitioning from a large language model company to a platform company for the AI era. The core logic is that platform value equals intelligence density multiplied by token throughput; when both dimensions are sufficiently strong, platform value emerges naturally.

At the product level, the M2.5 model has achieved globally leading performance in multiple productivity benchmarks. Alongside enhanced model capabilities, the average daily token consumption for the M2 model series in February 2026 was more than six times the level recorded in December 2025, validating market acceptance of its cost-effective approach.

Key takeaways from the earnings call include:

1. **Model Iteration Speed and Commercial Validation:** The company completed three version iterations from M2.0 to M2.5 within 108 days. Daily token consumption for the M2 series grew over sixfold since December 2025, confirming market acceptance of its high-value strategy. 2. **Progress in Multimodal Strategy:** Multimodal fusion is established as the inevitable path toward AGI. MiniMax has completed independent refinement of individual modalities and plans to launch the M3 series in the first half of 2026, showcasing synergistic advancements. Video generation has become the third-largest segment by API calls, with multimodal capabilities seen as a core barrier for capturing this market. 3. **Assessment and Focus on Agent Evolution:** The company confirmed the arrival of L3-level agents, stating the distinction between L4 and L5 lies in "single tasks" versus "multi-agent collaboration." Programming scenarios were validated first, but the potential market in office productivity (data analysis, document writing, PPT creation) is considered far larger. 4. **Differentiated Competitive Strategy:** The strategy involves focused investments. In 2023, the company abandoned development of a generic mobile personal assistant to concentrate resources on differentiated products like its editor and Conch Video. The R&D strategy prioritizes speed and excellence in specific capabilities over aiming for across-the-board leadership. 5. **Underlying Logic of R&D Efficiency:** The essence of AI competition is emphasized as model iteration speed and marginal efficiency, not simply burning money and resources. Training all modalities under a unified architecture is stated to be far cheaper than building separate systems, with synergies being continuously validated. 6. **Spillover Effects from Ecosystem Development:** The company's models are generating ecosystem spillover, from contributions to the Google Cloud ecosystem to leading call volumes on developer platforms like OpenRouter. Future efforts will focus on lowering usage barriers through product-level multimodal capabilities to build a more complete platform ecosystem.

**Revenue Acceleration with International Markets Contributing Over 70%**

Deconstructing the $79 million annual revenue, both major business segments experienced high growth. The open platform for enterprises and individual developers contributed approximately $26 million, a 198% year-on-year increase. Consumer-facing AI products, including MiniMax Agent, Conch AI, Talkie, and Xingye, contributed approximately $53 million, growing 143% year-on-year.

Internationalization is a notable feature of the revenue structure. In 2025, international market revenue accounted for over 70% of total revenue, with international revenue from the open platform also exceeding 50%. As of December 31, 2025, MiniMax has cumulatively served over 236 million users across more than 200 countries and regions, alongside 214,000 enterprise clients and developers from over 100 countries.

Expense trends indicate the initial emergence of scale effects. Sales and marketing expenses decreased by 40% year-on-year, while R&D expenditure increased by 33.8%, a growth rate significantly lower than revenue growth. Commercial momentum strengthened further entering 2026, with Yan Junjie revealing that new user registrations on the open platform in February 2026 were over four times the level of December 2025.

**Accelerated Model Matrix Iteration, M2.5 Sets New Programming Benchmark**

The call highlighted MiniMax's rapid model iteration capability. The company intensively launched three large language models—M2, M2.1, and M2-her—in Q4 2025, completing three generations of evolution from M2 to M2.5 in just 108 days.

The M2.5 model, released in February 2026, achieved globally leading performance in productivity scenarios. In programming, it set a new industry record on the SWE-bench Verified benchmark, showing a 37% efficiency improvement over the previous M2.1 generation. Breakthroughs were also achieved on cost—running a complex Agent at 100 tokens per second costs only $1 per hour. The company estimates a $10,000 budget could sustain an Agent running continuously for a full year. Since its release, M2.5 quickly ascended to the top of the Open Router rankings.

Multimodal capabilities advanced concurrently. The Conch 2.3 video model, Speech 2.6 voice model, and Music 2.0/2.5 music models were successively launched. By the end of 2025, the video models had helped creators generate over 600 million videos cumulatively, while the voice models had generated over 200 million hours of speech content.

Inference efficiency saw significant optimization. By February 2026, the computational cost per million tokens for the M2.5 series had decreased by over 50% compared to December 2025, while inference latency for the Conch video generation model was reduced by over 30%.

**Ecosystem Expansion Accelerates, Integration with Major Cloud Platforms and Toolchains**

MiniMax achieved a series of key developments in its commercial ecosystem. Global mainstream cloud platforms are accelerating the integration of its models—Google Vertex AI, Azure AI Foundry, Fireworks AI, and NetViews AI have all deployed MiniMax models. In the programming tools space, MiniMax has become the default model for mainstream platforms like OpenCode and Kilo Code.

In early 2026, Notion announced the integration of the M2.5 model, making it the first and only open-source model option on the platform. Yan Junjie stated this signifies a deepening of MiniMax's penetration into productivity scenarios.

Synergies with the OpenClaw project also released ecosystem effects. Yan Junjie mentioned that OpenClaw's founder, Peter, had previously publicly stated that M2.1 was his preferred best open-source model. MiniMax subsequently launched MaxClaw, further lowering the user barrier and promoting widespread model adoption within the developer community.

**Organizational AI Transformation Accelerates, Internal Agents Support 90% of Employees**

Regarding organizational transformation, Founder and CEO Yan Junjie disclosed that the company's internal Agent Interns now provide support to nearly 90% of employees, covering scenarios like software development, data analysis, operations management, talent recruitment, and sales marketing. He defined this practice as a core source of the company's competitive advantage.

Large-scale internal Agent deployment is yielding dual benefits. On one hand, the feedback loop between model iteration and product innovation is significantly accelerated. On the other hand, real-world deployment clearly exposes the shortcomings of current model capabilities, directly guiding the R&D priorities for the next generation. Yan Junjie observed a clear shift within the company, where employees are moving from "teaching the Agent how to work" to "observing how the Agent works."

**2026 Outlook: Betting on M3 Full-Modal Model, Transitioning to a Platform Company**

In the 2026 outlook, Founder and CEO Yan Junjie presented three core predictions: the software development field will witness a leap to L4-L5 level intelligence, with AI evolving from a tool to a colleague-level collaborator; workplace productivity scenarios will replicate the rapid penetration path seen last year in programming; and multimodal content creation will advance to directly generating medium-to-long-form production-ready content, with output formats increasingly resembling real-time streaming. He anticipates these trends will drive platform token demand to grow by one to two orders of magnitude.

To meet this demand, the company's next-generation flagship products, the M3 and Conch 3 model series, have been architecturally designed for these scenarios, with plans to launch integrated multimodal capabilities in the second half of 2026. Yan Junjie stated that MiniMax is one of only three Chinese companies leading in every modality and one of the few independent companies capable of executing simultaneously at both the product and model layers.

Strategically, Yan Junjie redefined a platform company in the AI era as one capable of defining and advancing new intelligence paradigms while continuously capturing the commercial value created by these paradigm shifts. This definition clearly distinguishes it from the internet-era platform paradigm centered on traffic gateways.

Management stated that MiniMax's goal is to become a platform company for the AI era, with core drivers being the continuous enhancement of model capabilities and the deep mining of customer value.

In terms of strategic execution, the company remains focused on the keywords "full-modal" and "high-quality," making deliberate choices. Yan Junjie revealed that as early as 2023, the company decided against developing a generic mobile personal assistant, judging it difficult to create unique value there. Instead, resources were concentrated on differentiated products like its editor and Conch Video.