No longer content with being merely another large model developer, MiniMax now wants to define what an artificial intelligence platform company looks like as the sector enters its next phase.
On March 2, MiniMax released its annual results for the year ended December 31, 2025, marking its first earnings report since listing on the Hong Kong Stock Exchange on January 9. Since its debut, its share price has risen more than threefold.
The company generated USD 79 million in revenue in 2025, up 158.9% year-on-year (YoY). More than 70% of that revenue came from international markets, underscoring its expanding global footprint.
MiniMax has established a dual revenue structure spanning both B2C and B2B markets. In 2025, revenue from its AI-native products rose 143.4% YoY to USD 53.1 million, driven primarily by word-of-mouth growth for products such as Hailuo AI and Talkie, as well as increased user willingness to pay. Revenue from its open platform and enterprise services grew 197.8% to USD 26 million, reflecting broader adoption of its model APIs in enterprise applications.
Profitability metrics improved, though the company remains in the red. In 2025, MiniMax reported gross profit of USD 20.1 million, up 437.2% YoY. Gross margin expanded 13.2 percentage points to 25.4%. Adjusted net loss totaled USD 250 million, with the loss ratio narrowing substantially, indicating improved commercialization efficiency.
Its cost structure also shifted. R&D expenses increased 33.8% YoY, while marketing expenses declined 40.3%, suggesting tighter spending discipline alongside continued technical investment.
Operationally, both AI-native products and open platform revenue recorded rapid growth. MiniMax said it now serves more than 236 million users across over 200 countries and regions, as well as 214,000 enterprise customers and developers in more than 100 countries and regions.
On the technical front, MiniMax earlier this year released its next-generation foundation model, M2.5. The company said the model achieved globally leading performance in programming and agent-based scenarios. Model usage rose accordingly. By February, average daily token consumption for the M2 series text models had increased to more than six times the level recorded in December 2025.
During the earnings call, management disclosed that annual recurring revenue (ARR) surpassed USD 150 million in February.
Founder and CEO Yan Junjie said the company had strengthened its multimodal capabilities across language, video, voice, and music, while accelerating international expansion. He added that anticipated gains in programming, office productivity, and multimodal creation could significantly increase token demand.
Yan added that MiniMax intends to transition from a foundation model developer to an AI platform company focused on scaling intelligence density, infrastructure, and global commercialization.
From model developer to AI platform company
During the earnings call, Yan elaborated on the rationale behind the company’s transition.
He proposed assessing AI platform value by examining the intersection of intelligence density and token throughput. In his view, competition in the sector will hinge on whether companies can expand the boundaries of machine intelligence. As those boundaries expand, new use cases and user segments emerge, creating additional ecosystems and commercialization opportunities.
MiniMax outlined three forecasts for industry trends in 2026.
- In programming, Yan said Level 4 or 5 intelligence could begin to emerge, shifting AI systems from tools to “colleague-like” collaborators.
- In office productivity, he said the sector could mirror last year’s advances in programming, with AI agents improving delivery capability and penetration in workplace scenarios.
- In multimodal creation, MiniMax expects AI systems to generate directly deliverable mid- to long-form content and move toward streaming, real-time outputs.
Yan said the convergence of these trends could increase token volume by one to two orders of magnitude.
MiniMax has begun developing its M3 and Hailuo 3 series models while continuing to optimize inference architecture and compute efficiency. Internally, it reported that AI agents now support more than 90% of employees’ daily work scenarios. The company said these deployments create feedback loops for model iteration and help refine R&D priorities.
Taken together, 158.9% revenue growth and ARR exceeding USD 150 million provide early signals supporting its platform ambitions. The next challenge will be sustaining efficiency gains while further increasing intelligence density.
Three model generations in 108 days
Between the fourth quarter of 2025 and early 2026, MiniMax upgraded its language models from M2 to M2.1, and then to M2.5, over 108 days. The company said this pace matches or exceeds that of global peers such as Anthropic, OpenAI, and Google.
Released in February, M2.5 improved programming efficiency by 37% compared with M2.1, according to MiniMax. It also said the model made complex agent operations economically scalable.
From M2 to M2.5, both performance and usage increased. By February, average daily token consumption for the M2 series text models had risen to more than six times the December 2025 level. Token consumption in its Coding Plan scenario increased more than tenfold.
As model iteration cycles shorten, benchmark rankings may lose relevance quickly. The competitive focus is shifting from absolute capability at a given moment to the speed of improvement.
MiniMax attributed its iteration pace to optimization of its mixture-of-experts architecture and its native agent reinforcement learning framework, Forge, which it said delivered gains in both performance and efficiency.
Beyond text, the company has expanded into video, voice, and music.
Its Hailuo 2.3 video model, released in October 2025, improved character motion, image quality, and stylization, according to the company. A “Fast” variant reduced batch creation costs by up to 50%. An upgraded media agent within the Hailuo AI product supports full-modality creation, enabling users to generate finished videos from a single prompt. By the end of 2025, MiniMax said its video models had been used to generate more than 600 million videos globally.
Its Speech 2.6 voice model, also released in 2025, focused on real-time voice agent scenarios, achieving ultra low latency and supporting more than 40 languages. By year’s end, MiniMax said its voice models had generated more than 200 million cumulative hours of audio worldwide.
Its Music 2.0 and 2.5 models improved control over vocal styles and emotional expression, producing compositions up to five minutes in length, according to the company.
MiniMax also launched MaxClaw, a cloud-based AI assistant built on OpenClaw. The service does not require local deployment or additional API fees and operates continuously in the cloud. Users with a basic MiniMax Agent subscription can access it immediately. The company said the product gained traction shortly after launch.
Over time, MaxClaw could strengthen user retention and willingness to pay by delivering ready-to-use automated services. Increased usage may also generate feedback data that supports model improvement, reinforcing a cycle between product adoption and model iteration.
Endorsement from global investment banks
In recent months, several global and domestic investment banks, including JPMorgan Chase, UBS, Goldman Sachs, CICC, CSC Financial, and Guotai Haitong, initiated coverage on MiniMax. They assigned buy, overweight, or outperform ratings, with target prices above HKD 1,000 (USD 128).
JPMorgan said MiniMax stands out in the foundation model segment for combining technical capabilities, multimodal commercialization potential, and global scalability. It cited the company’s benchmark performance, diversified B2B and B2C monetization pathways, and international expansion strategy.
UBS said MiniMax is positioned to capture AI growth opportunities in both domestic and overseas markets, pointing to its multimodal strategy, R&D execution, and global commercialization footprint.
Goldman Sachs highlighted its diversified revenue structure and multimodal product portfolio.
CICC described MiniMax as one of the few Chinese companies to demonstrate both foundation model capability and global commercialization of AI-native applications.
CSC Financial characterized foundation model competition as an evolving global race, arguing that MiniMax’s sustained model development, cost efficiency, overseas expansion, and open-source and agent ecosystem contribute to its competitive position.
Guotai Haitong cited improvements in technical performance and efficiency, cost control that supports commercialization, and momentum from product iteration and ecosystem expansion in the emerging agent landscape.
JPMorgan added that China’s AI sector is moving beyond an early phase marked by a proliferation of large model startups into a stage defined by commercialization capability, model originality, and global expansion. In that context, it said MiniMax represents not only a technology company, but also China’s broader participation in the global AI value chain.
This article was adapted based on a feature originally written by Stone Jin and published on IPO Zaozhidao. KrASIA is authorized to translate, adapt, and publish its contents.