MiniMax Conference Call: Focusing on "Full Modality" and "High Quality," moving beyond simple "model competition," evolving towards an AI platform ecosystem.

```

Chinese AI unicorn MiniMax released its first annual financial report since going public, delivering an unexpectedly strong growth performance.

Full-year revenue for 2025 reached $79.04 million, up 159% year-on-year and about 10.6% higher than Bloomberg’s compiled market expectation of $71.4 million. Annual recurring revenue as of February 2026 surpassed $150 million, indicating significant acceleration of commercialization. Full-year gross profit jumped 437% year-on-year to $20 million, with gross margin rising from 12.2% in 2024 to 25.4%. Full-year adjusted net loss was $250 million, basically flat year-on-year.

During the post-earnings conference call, founder and CEO Yan Junjie elaborated on the strategic evolution direction: MiniMax is transitioning from a large model company to a platform company for the AI era. Its core logic is that platform value equals intelligence density times token throughput—when both dimensions are strong enough, platform value will naturally emerge.

On the product front, the M2.5 model has achieved global-leading performance in multiple productivity benchmarks. With improved model capability, average daily token consumption for M2 series models in February 2026 was more than six times the December 2025 level, verifying market acceptance of the high cost-performance route.

Key points of the conference call:

Model iteration speed and commercialization validation: Three iterations from version 2.0 to 2.5 completed in the past 108 days; daily token consumption of the M2 series models has grown over sixfold since December 2025, validating market demand for high cost-performance.Stage results of multimodal strategy: Multimodal integration is recognized as the inevitable path to AGI; independent refinement of each modality has been completed; M3 series models expected to launch in first half of 2026, showing coordinated evolution across modalities. Video generation has become the third-largest API call market, and multimodal capabilities are seen as the core barrier to dominate that market.Agent evolution: judgment and layout: L3-level agents have clearly arrived; L4 and L5 divide between "single-task" and "multi-agent collaboration". Programming scenarios validated first, but potential market for workplace scenarios (data analysis, document writing, PPT production) considered much larger than programming.Differentiated competition strategy: Strategic focus and restraint; abandoned mobile general-purpose personal assistants in 2023, concentrating on editor, Hailuo Video, etc., to build differentiated products. R&D strategy does not pursue full leadership, instead opens market with "fast iteration" and "specific outstanding capabilities".Underlying logic of R&D efficiency: Emphasizes that essence of AI competition is not burning cash or resources, but model iteration speed and marginal efficiency. Cost of unified architecture for full-modal training is much lower than separate systems, and synergy has been repeatedly validated.Ecological spillover effects: Model has generated excess value at the ecosystem level—from contributions to Google Cloud ecosystem to being top-called model on developer platforms like OpenRouter; in future, product-level multimodal abilities will further lower usage barrier, building a more complete platform ecology.

Revenue acceleration: international markets contribute over 70%

Breaking down MiniMax's annual $79 million revenue, both major business segments saw rapid growth. The open platform for enterprises and individual developers contributed about $26 million, up 198% year-on-year; consumer-facing AI products, including MiniMax Agent, Hailuo AI, Talkie, and Xingye, contributed about $53 million, up 143% year-on-year.

Internationalization has become a notable feature in company’s revenue structure. In 2025, revenue from international markets accounted for over 70% of total revenue, international revenue from the open platform was over 50%. As of December 31, 2025, MiniMax had served a cumulative 236 million users across more than 200 countries and regions, and 214,000 enterprise clients and developers from over 100 countries and regions.

The expense side shows scale effects starting to emerge. Sales and marketing expenses fell 40% year-on-year, R&D expenditure grew 33.8% yet far slower than revenue growth. Into 2026, commercialization momentum further strengthened—Yan Junjie revealed that new user registrations on the open platform in February 2026 were over four times the December 2025 level.

Faster model matrix iteration, M2.5 resets programming benchmarks

On the technical side, MiniMax exhibited rapid model iteration capability. In Q4 2025, M2, M2.1, and M2-her large language models launched in quick succession, with three generations evolving from M2 to M2.5 in just 108 days.

M2.5, released in February 2026, achieved global-leading performance in productivity scenarios. In programming, it set a new industry record in the SWE-bench Verified benchmark, with 37% efficiency improvement over the previous M2.1. Cost breakthroughs were made—at 100 tokens output per second, M2.5 runs complex agents at just $1 per hour. The company estimates that a $10,000 budget can support year-round agent operation. Since its release, M2.5 quickly topped the Open Router leaderboard.

Multimodal capabilities are advancing in parallel. Hailuo 2.3 video model, Speech 2.6 voice model, and Music 2.0/2.5 music models have landed. By end-2025, video model had helped creators generate over 600 million videos; voice model generated over 200 million hours of voice content.

Inference efficiency optimization achieved impressive results. By February 2026, M2.5 series inference computational cost per million tokens was over 50% lower than December 2025, and Hailuo video generation inference latency was reduced by over 30%.

Eco layout speeds up: Leading cloud platforms and toolchains integrate one after another

MiniMax has achieved key business ecosystem milestones. Leading global cloud platforms are swiftly incorporating its model capabilities—Google Vertex AI, Azure AI Foundry, Fireworks AI, and NetViews AI have all deployed MiniMax models. In programming tools, MiniMax has become the default model for mainstream platforms like OpenCode and Kilo Code.

In early 2026, Notion announced M2.5 integration, making it the platform’s first and only open-source model option. Yan Junjie said, This marks further penetration of MiniMax into productivity scenarios.

Collaboration with the OpenClaw project also released ecological effects. Yan Junjie mentioned OpenClaw founder Peter previously publicly stated M2.1 is his preferred best open-source model. MiniMax quickly launched MaxClaw, further lowering user adoption barriers and driving broad uptake in the developer community.

AI transformation speeds up internally; Agents cover 90% of staff

On organizational transformation, founder and CEO Yan Junjie revealed, internal agent interns now support nearly 90% of employees, covering software development, data analysis, operations management, talent recruitment, sales and marketing. He defined this practice as a core source of company competitive advantage.

Large-scale internal agent deployment brings dual benefits: feedback loop between model iteration and product innovation is significantly accelerated; actual deployment environment clearly exposes model shortcomings, directly guiding next-generation model R&D priorities. Yan Junjie observed, the company is experiencing a clear shift from “teaching agents how to work” to “observing how agents work”.

Outlook for 2026: Betting on M3 full-modal model, leaping to platform company

In the 2026 outlook, founder and CEO Yan Junjie presented three core views: software development will see a leap to L4 to L5 level intelligence, AI will evolve from tool to colleague-level collaborator; workplace productivity scenarios will replicate last year’s rapid penetration path seen in programming; multimodal content creation will move into direct generation of mid-to-long format production-grade content, with output ever closer to streaming real-time. He expects these trends will drive platform token demand up one or two orders of magnitude.

To meet this demand, next-generation flagship M3 and Hailuo 3 models have been architected for these scenarios, planning to launch multimodal fusion capabilities in second half of 2026. Yan Junjie claims, MiniMax is one of only three Chinese companies to achieve leading performance in every modality, and among the few able to execute at both product and model layers independently.

Strategically, Yan Junjie redefines platform companies for the AI era as: those able to define and drive new intelligent paradigms, continually capture business value created by paradigm shifts—distinct from Internet-era platforms centered on traffic entry points.

Management states, MiniMax aims to be a platform enterprise for the AI era, driven by continual enhancement of model capabilities and deep exploration of customer value.

Strategic execution-wise, The company insists on acting around the two key words “full modality” and “high quality”, knowing what to do and not do. Yan Junjie revealed that in 2023, the company decided not to make a mobile general-purpose personal intelligent assistant, judging that it could not create unique value; instead resources were concentrated on differentiated products like editors and Hailuo Video.

The following is the full conference call transcript (AI-generated):

2025 Annual Earnings Conference Call

Operator:
Hello ladies and gentlemen, thank you for waiting. Welcome to MiniMax’s 2025 full-year financial results conference call. Please note that management’s keynote and Q&A will have English simultaneous interpretation.

The English line will be listen-only mode. I now turn the line over to Ms. Yu Meiqi, Director of Investor Relations at MiniMax.

Unnamed speaker:
Thank you, operator. Good evening, good morning everyone. Welcome to MiniMax’s 2025 annual financial results conference call. Before we begin, please note today’s discussion may contain forward-looking statements involving risks and uncertainties. Actual results may differ. Except as required by law, the company does not undertake to update any forward-looking information.

For important information about this call, including forward-looking statements, please see public disclosures released earlier or the 2025 full-year earnings announcement and financial status as of December 31, 2025. Today, management will also discuss certain non-IFRS financial indicators that are supplementary and should not substitute for IFRS-based results. For definitions, reconciliations and risk factors of non-IFRS indicators, please see our 2025 earnings announcement.

Management will speak mainly in Chinese. Third-party interpreters will provide English simultaneous translation for the keynote and Q&A. Please note, English translation is for convenience only; in case of any ambiguity, the original language from management prevails. Unless otherwise specified, all monetary figures are in USD. I now turn the call to our Founder and CEO Dr. Yan Junjie.

Yan Junjie, Founder, Chairman, CEO and CTO:
Dear investors and analysts, good evening. This is Yan Junjie. Thank you for joining our first earnings call since IPO. I’d like to share our progress over the past year and strategic focus for the next phase of growth.

First, a review of 2025. For MiniMax, the theme was "solid foundation". In 2025, we established full-modal R&D capability, owning globally competitive models in key modalities: language, video, voice, and music. We upgraded our products through continuous innovation—including developer and enterprise-focused open platform, and consumer-oriented MiniMax Agent, Hailuo AI, Xingye, etc. We further deepened our global presence. In Q4, we released three updated large language models: M2, M2.1 and M2-her. M2 redefined the balance of performance, cost and speed, integrating three key abilities: programming, tool usage, and deep search.

Performance approached globally leading levels. After launch, M2 was rapidly adopted by the global developer community, becoming the first Chinese model on OpenRouter to exceed 50 billion tokens of daily consumption, and topped Hugging Face global trends. Based on M2, we quickly released M2.1, focused on improving performance in complex real-world tasks—especially programming and workplace scenarios, with better comprehension and execution of multi-step instructions. M2-her serves as the foundational model for our AI interaction product, Xingye.

It’s designed to provide more natural, personalized conversation and ranked first globally in long-context dialogue tests. In February, we launched M2.5, delivering globally leading performance in key productivity scenarios including programming, tool usage, and workplace applications. In programming, M2.5 set a new industry record in FWE Bench Verified benchmark, achieving 37% efficiency improvement versus prior M2.1.

More importantly, M2.5 made complex agent operations economically feasible. At 100 tokens output per second, one hour continuous runtime costs only $1; $10,000 budget enables year-round agent operation. The breakthrough also quickly boosted usage—after launch, M2.5 topped OpenRouter’s charts.

From M2 to M2.1, now M2.5, each generation saw significant improvement in capability and adoption. In February 2026, daily token consumption for M2 series was over 6 times December 2025’s figure, programming-token demand grew over tenfold. In multimodal, we have now established comprehensive video, voice, and music model capabilities.

In October last year, we released video model Hailuo 2.3, making significant advances in character motion, visual quality, and style expression. We also launched a faster model, reducing batch creation cost by up to 50%, and upgraded media agents in Hailuo AI for one-click multimodal content creation.

By year-end 2025, our video model helped global creators generate over 600 million videos. In October, we launched voice model Speech 2.6, optimized for voice agents, boosting interaction and achieving global-leading ultra-low latency across 40+ languages. By year-end, our voice model had generated over 200 million hours of audio, becoming central infrastructure for voice AI ecosystem. Newly released Music 2.0/2.5 also made progress, reliably handling broad vocal styles and emotional expression.

In developing these models and products, we continued AI-native organizational evolution. Internally, our agent interns now support nearly 90% of employees, in software dev, data analysis, operations, talent, sales, etc. We act as a proving ground for AI-native organizational capability, steadily improving R&D efficiency. In January, we productized this capability, launching MiniMax Agent 2.0, enabling agents direct access to users’ local workspace. We also introduced expert agent creation for professional usecases.

By end-February, professionals had created over 50,000 expert agents to solve specialized challenges via deep knowledge/integration. Even before OpenClaw project gained broad attention, its founder Peter had praised MiniMax, calling M2.1 his preferred and best open-source model. After OpenClaw’s launch, M2 series’ performance/cost advantage enabled wider developer adoption at much lower cost. Our agent products actively supported OpenClaw, launching MaxClaw, further lowering user access barriers.

Next, let’s talk about commercialization. For the full year, 2025 revenue reached $79 million, up 159%; AI product revenue $53 million, up 143%; open platform revenue about $26 million, up 198%. See next slide.

2025 revenue accelerated. For developer and enterprise-focused open platform, new user registrations in February 2026 were over four times the December 2025 figure. As of December 31, 2025, we had served over 236 million users from 200+ countries/regions and 214,000 enterprise clients/developers from 100+ countries/regions. International market revenue accounted for over 70% of full-year total, and over 50% of open platform revenue.

Since M2.5’s launch, strong traction in international markets has built fast—new global client interest is robust, positive user word-of-mouth spreading. Leading global cloud providers and AI-native platforms—Google Vertex AI, Microsoft Azure AI Foundry, Fireworks AI, NetViews AI—all deployed MiniMax models. We also became default model for leading coding platforms like OpenCode and Kilo Code. Earlier today, Notion launched M2.5 as its first and only open-source model option.

Meanwhile, we drove meaningful improvement in computing efficiency and cost. Thanks to algorithm optimization, operator implementation and iterative improvements in code/decoding engineering, by February 2026, M2.5 per-million-token inference computation cost had dropped by over 50% versus December 2025. At the same time, high-fidelity video generation inference latency reduced by over 30%.

As model capability iterates and improves, new scale efficiencies have emerged. Full-year 2025 gross profit reached $20 million, up 437%, gross margin improved to 25.4% versus 12.2% last year. Expenses—sales/marketing expenses dropped 40% year-on-year, R&D up 33.8% but still far below revenue growth. 2025 adjusted net loss was $250 million.

As commercialization continues and model optimization brings cost benefits, we tightened adjusted net loss margin. For the first two months of 2026, strong growth momentum is evident. By February 2026, annual recurring revenue passed $150 million. Next, I’ll share our outlook for the future.

We believe intelligence levels will rise sharply in 2026. Our efforts focus on three areas. First, software development will see emergence of L4/L5-level intelligence, marking AI’s shift from tool to colleague collaborator. Second, in professional work scenarios, similar progress speed as last year’s programming field will be seen.

AI agents’ delivery and penetration in workplace scenarios will surge. Third, multimodal creation will move toward generating immediately usable mid-long form content and real-time streaming formats. All these developments signal technical challenges, huge expansion in intelligent supply, and massive innovation windows at application layer. They mean demand for our platform will rise sharply, token quantity potentially growing by one or two orders of magnitude.

Our next-gen M3 and Hailuo 3 models are designed for such demands. At the same time, we are rapidly strengthening infrastructure, continuously attracting top talent, shifting focus from training efficiency optimization to faster R&D and iteration. Strategically, we are evolving from a large model company to an AI-era platform company. In the Internet era, platforms were traffic gateways.

But in the AI era, platforms are those that define and advance new intelligence paradigms and capture the product/commercial value arising from paradigm shifts. This requires ability to shape new intelligence frameworks, continuous innovation in tech and products, scalable infrastructure, and high-efficiency token throughput. We believe we are one of the few companies already building and strengthening these abilities. Therefore, an AI platform company’s value can be simply described as: intelligence density provided times token throughput. When both are strong, platform value naturally follows. Amid this historic industry inflection, our capability comes from two sources. AI industry is accelerating rapidly. Model breakthroughs, agent deployment and mature monetization mode are spreading industry-wide.

Strong growth momentum is evident. We are confident we can be a core constructor of AI platform ecosystems. That’s all for our prepared remarks. Now we’re ready for your questions.

Q&A

Operator:
Q&A session

Operator:
We begin Q&A. (instructions) Your first question is from Morgan Stanley’s Gary Yu.

Gary Yu:
Thank you management. Appreciate your remarks.
Your vision is to become an AI platform company. How do you define a platform company in the AI era? Why do you think a startup like MiniMax can become one? Thank you.

Yan Junjie, Founder, Chairman, CEO and CTO:

Thank you for your question. This is something we discuss and think about internally on a long-term basis. As we said, when intelligence boundaries are pushed, it produces many new scenarios, customers and users, forming a new ecosystem and new commercial dividends, as seen with coding or visual/image generation. Why does MiniMax have a chance to become an AI platform company? Several reasons. First, the AI market isn’t zero-sum.

Annual incremental market is larger than existing stock market. It’s not winner-takes-all; as long as you have unique, differentiated innovation, you can find your market fit. We believe in the next two or three years, our model R&D and infrastructure capabilities can create new scenarios, with huge innovation potential in coding, office efficiency and interactive entertainment.

In this high-growth, rapidly changing market, opportunity exists at three levels. First, model layer—a key factor is reliance on sustained accumulation and faster iteration. For example, in 108 days we released M2, M2.1, M2.5, each driving quick user/API usage growth. From day one, we've accumulated cross-modal capability; we’re the only company in China doing so, which gives us a favorable position for the inevitable multimodal trend. Second, product layer. MiniMax is China’s first company to focus on both product and model.

“Model+product” forms a stronger entry barrier. Treating the model as a product is hard for competitors to copy. Third, ecosystem layer. We use differentiated ability to build an open system, as seen in the OpenClaw ecosystem, which uses many of our models for development. Our models suit high-throughput product scenarios. Via further integration, we also lower user adoption barrier, hence a lot of code contribution. We can help the ecosystem grow quickly. Going forward, this is just the start of our internal ecosystem building. Next, we’ll focus on creating the next-gen full-modal M3 series, establishing clear model differentiation.
On the other hand, we want to build a unique product/ecosystem around the intelligence we provide. Outside a few big players, we believe we are Asia’s only company innovating at both model and product layers. Thank you.

Yan Junjie, Founder, Chairman, CEO and CTO:
Next question, thank you.

Operator:
Next question from JPMorgan’s Alex Vovk. Please go ahead.

Alexander Vovk:
Thank you for your time. Congrats on strong results.
I want to ask about multimodal, which you call AI’s endgame. If competitors perfect single-modal first, then move to cross-modal, they might move faster than you. Does your early cross-modal focus slow you down, is it a burden?

Yan Junjie, Founder, Chairman, CEO and CTO:
Thank you. This is a question we've faced since day one. Let me explain why we focus on cross-modal. We see multimodal integration as the fundamental premise for sustained intelligence improvement. In the past six months, models broke through via multimodal fusion, as seen in Nano Banana Pro, which combines visual understanding/generation and expands intelligence boundaries. We use a two-stage method. We're now entering stage two, after four years of stage one.

We steadily built industry-leading models in each modality, earning positive reputation and market recognition. Many models in each modality have been delivered, achieving notable results. Next, key is integration/fusion to achieve bigger breakthroughs—M3 model this year is aimed at that. Two points: first, each modality accumulation is a long process.

Data to single-modal, then to fusion, needs considerable time. This is the base for long-term capability and our differentiation. We're one of only three companies in China with leading models in every modality. Second, video generation (besides coding/agent tasks) is the biggest market. We believe we’ll see near real-time mid-long format content, and are confident in realizing it. As you say, does our strategy slow R&D? Challenges exist but are inevitable.

Since founding, AGI is multimodal input/output. We designed organizational structure so cross-modal basic capabilities are reusable. As seen, in this AI-native structure, building full-modal is not more costly for us than for other startups, and is much less than big tech companies. Each modality has competitive models.

In some cases, even outperforming single-modal focused companies. Our technical vision and positioning is validated repeatedly, and will become even clearer going forward. Thank you.

Unnamed speaker:
Next question, thank you.

Operator:
Next question from UBS's Kenny Fong.

Kenneth Fong:
Congrats on strong IPO results. You mentioned L4-L5 programming intelligence is arriving, and many claim software companies may be replaced by agents. How should we view this change? Where are you positioned?

Yan Junjie, Founder, Chairman, CEO and CTO:
That's very important.
Let me explain L4-L5 intelligence, the future of programming intelligence, and our role. L3 are agents in use today; L4/L5 are colleague and organizational-level intelligence. For example, to build world-leading models, many people collaborate on algorithms, experiments, optimization, data processing and ops—a massive workload. We think L4 can handle many innovative tasks, such as running experiments based on a research paper and proposing solutions for many challenges (achieving innovation). L5-level intelligence needs not just one person, but collaborative teams.

Programming is just one part of agent intelligence—the first productivity capability to be validated. Beyond programming, workplace productivity will, in the coming year, replicate the swift progress seen last year in programming. The market is growing, and I think it's bigger than programming. How do we see ourselves? I think there's a huge market ahead. Programming models help more people write code, and write it better; but programmers are a small portion of the workforce.

Most of the workplace uses non-code software (data analysis, financial modeling, slide deck creation for earnings calls, etc.), representing a much larger market than programming. We've made initial progress in programming and agents, occupying a distinct market position with minimal resources.

The bigger market penetration is just beginning. We move fast; as I said, evolution from M2 to M2.3 took only 108 days. We maintain industry-leading iteration speed, each generation sees notable improvements in capability and adoption—highlighting our R&D and scale capacity.

We built M2 with limited resources, which are now increasing. With faster model improvement, better models further the ceiling. Our past performance was based on M2; we expect M3 to unlock greater potential, forming a positive flywheel effect.

Beyond speed, we create differentiated models, validated repeatedly in recent months. As I said, the market is huge, and technology paths will diverge. We need to know if we can define the tech roadmap—not win at every dimension, but define model abilities showing our unique edge. For M2, Hailuo 2 and Voice 2, each has clear differentiation and quickly rallies market interest; features include low latency and strong cost-effectiveness.

These characteristics set us apart and help us win more market share. With expanded organization and resources, our deep understanding of model evolution and tech paths will further boost differentiation and value. In sum, we are confident in growing our share via programming agents and broader productivity markets, achieving new breakthroughs—with faster iteration and stronger differentiation positioning for greater market success. Thank you.

Operator:
Next question from Goldman Sachs.

Analyst:
Thank you for your comments. We know industry includes tech giants, startups, and open models.
Where do you compete? What are your priorities?

Yan Junjie, Founder, Chairman, CEO and CTO:
As mentioned, we're building and aiming to be an AI-era platform company, driven by continuous intelligence density improvement and scalable commercial growth. Compared to other AI firms, we differ in several ways. First, strategic positioning: from day one, we focused on full-modal models to boost intelligence density and expand boundaries, creating differentiated value.

Simultaneously, we build scalable products/businesses around model intelligence density, concentrating resources where differential value can be achieved. For example, in 2023, we opted not to build a generic mobile assistant (like Doubao or ChatGPT), as we did not see unique value in that area. Instead, we focused on differentiated model R&D and product innovation, not burning cash. Hailuo and MiniMax Agent products are our focus.

This strategic choice strengthened differentiation and improved our odds. Another example—we committed to cross-modal foundational model development from day one. As mentioned, modality accumulation is crucial.

We've now reached the key stage for cross-modal fusion, placing us favorably for inevitable full-modal trends. Second, R&D efficiency. In the AI era, ultimate success isn’t about how much money/resources you burn, but about the speed of intelligence improvement.

This speed comes from R&D efficiency, which translates into larger market share and higher efficiency. We emphasize and practice this, applying it to every R&D stage (algorithm, experiment design, iteration cycles, etc.). We leverage our flexible organizational structure combining top-down and bottom-up approaches, cross-modal reuse of experience/infrastructure.

This ensures we stay ahead. Over the long term, we believe only a few global AI platform products will naturally lead the industry. We are one of the few independent firms with both significant advantages and clear differentiation to win in competition.

Operator:
Next question from CITIC Securities’s Yu Zhonghai.

Analyst:
Thank you. Congrats on strong results. You mentioned in the first two months of 2026, token consumption for M2 series was already 6x last December’s figure.
Is this explosive growth a one-off windfall, or the start of a sustainable trend? Because OpenClaw token demand has surged, so I'm asking. Is this temporary, or the beginning of a long-term trend?

Yan Junjie, Founder, Chairman, CEO and CTO:
Thank you. We believe this is the start of a sustainable trend—not a one-off windfall. Industry growth is often a step function, not linear. We keep releasing new models, seizing industry opportunities. At the core is our R&D strategy—preparing resources/capabilities ahead, and defining each generation based on our understanding of intelligence evolution. Beyond M2, next wave of growth is supported by several factors. In fact, since H2 2025, we have been actively preparing capacity for multiple high-impact productivity opportunities appearing in 2026.

Growth will become increasingly diversified. Programming has huge room to grow. It's already quite good as an assistant tool, and we think will advance further, evolving from tool assistant to colleague collaborator, and up to more advanced operator intelligence.

Based on tech reserves, R&D progress/judgment, we believe this is likely to unfold this year. Second, workplace scenarios are broader/larger than programming, involving many professions, tools, and complex problems—and many tasks can't be validated conventionally; we're actively preparing for such challenges. We expect workplace will see rapid progress similar to programming. In the multimodal realm, we believe we’ll greatly lower adoption barriers and make better models to generate directly usable, longer videos.

Model competition is about winning/losing; every company faces this. No one can guarantee perpetual leadership, but we are confident that we will keep winning in crucial areas. A key strategy—push technical boundaries, use breakthroughs to build a bigger ecosystem by products/models. The ultimate goal is to capture dividends. We are confident we’ll grow with the industry and expand our capabilities, R&D efficiency, product innovation and global monetization into lasting organizational advantage.

Operator:
Next question from Jefferies’ Thomas John.

Thomas John:
Good evening. Thanks for answering. You mentioned internal agent interns now cover nearly 90% of employees.
What insights has this brought? How does it feed back into your product/tech development?

Yan Junjie, Founder, Chairman, CEO and CTO:
Thank you. We're not just an AI company. Our goal is to build a truly AI-native platform company. Along with AI model research, we want to become an AI-native company.

This is a critical organizational goal. We focus on two things: first, speed—progress pace. As a startup with limited resources, we need to maximize efficiency to survive and succeed, so we've always used AI agents internally, many employees use them daily.

We've observed a clear trend. In many cases, we're shifting from people teaching agents how to work, to watching agents work—sometimes, agents surprise us. This not only shortens workflow, but lets every process benefit from improved intelligence.

From model iteration, product innovation, to customer service, our feedback and iteration speed are increasing. Employees can focus more on higher-value work, accelerating organizational thinking and innovation. It also feeds back into model R&D, letting us define target intelligence: as agents deploy in the organization, we observe that even the best model today still makes mistakes or can't complete tasks.

These gaps expose highest economic value; they guide priorities for next-generation model/agent R&D. The more agents we deploy, the clearer the direction for model iteration.

In the past few months, our model iteration speed, revenue growth, customer service abilities, and token throughput have all improved. This lets us define new model goals faster. We’re maximizing AI value internally, and as we said, we've seen positive flywheel effects in building an AI-native company. It’ll become a critical organizational competitive advantage. Thanks again for joining today. If you have further questions, please contact our IR team.
Thank you.

Risk warning and disclaimerThe market carries risk; investments should be made cautiously. This article does not constitute personal investment advice, nor does it consider individual users' special investment goals, financial circumstances, or needs. Users should determine whether the opinions, views, or conclusions herein fit their specific situation. Investment is at your own risk. ```