MiniMax

MiniMax Group Inc. (HKEX: 00100) develops foundation models across language, video, speech and music, delivering them through consumer products and an Open Platform for enterprises and developers. Its work spans long-context model architectures, reasoning, coding and agent workflows. In its unaudited first-half 2026 results, MiniMax reported US$116.6 million in revenue, with Open Platform and other AI-based enterprise services accounting for 63.4%.[5][6][7][8]

That platform business followed earlier work on long-context foundation models and a portfolio of conversational and media-generation products.[7][5]

MiniMax dates its founding to early 2022. Founder Yan Junjie previously spent more than six years at SenseTime, including as vice president and vice-head of its research institute. Yun Yeyi also came from SenseTime, where her roles covered fundraising, strategy and innovative business. MiniMax assigns her product and commercialization responsibilities; its management biographies identify Zhao Pengyu with language-model research and engineering and Zhou Yucong with visual-model research and engineering.[3][4]

On January 14, 2025, MiniMax and an alphabetically listed team of 89 co-authors published MiniMax-01: Scaling Foundation Models with Lightning Attention. The Text-01 and VL-01 series combined Lightning Attention with a mixture-of-experts architecture: 32 experts, 456 billion total parameters and 45.9 billion activated for each token. The team developed parallelization and computation–communication overlap techniques for training and inference across contexts spanning millions of tokens.[7]

The researchers reported that Text-01 supported up to one million tokens during training, extrapolating to four million during inference. VL-01 added vision-language capabilities through continued training on 512 billion vision-language tokens. These were distinct training and inference limits, rather than a single interchangeable context-window figure.[7]

MiniMax released M1 on June 16, 2025, making complete model weights and technical reports available. The hybrid-attention reasoning model used Lightning Attention and supported a one-million-token context and an 80,000-token reasoning output. Its reinforcement-learning algorithm, CISPO, clipped importance-sampling weights rather than using traditional token updates.[8]

MiniMax reported that M1’s reinforcement-learning phase used 512 H800 GPUs for three weeks at a rental cost of US$534,700; that figure covered the RL phase, not pretraining or total development costs. The release credited vLLM and Transformer with providing their own inference deployment support and separately described MiniMax’s collaboration with SGLang to advance deployment support.[8]

During the fourth quarter of 2025, MiniMax updated M2, M2.1 and M2-her. M2 emphasized coding, tool use and deep search; M2.1 focused on complex coding and workplace tasks involving multi-step instructions. M2-her supplied the underlying conversational model for Talkie and Xingye, separating interactive conversation from the productivity focus of the other M2 models.[5]

In October 2025, MiniMax released the video model Hailuo 2.3 and the speech model Speech 2.6. Hailuo AI’s Media Agent supported content creation from a simple description, while Speech 2.6 targeted voice-agent interactions and supported more than 40 languages. MiniMax reported that, by year-end, its video models had helped generate more than 600 million videos and its speech model more than 200 million hours of speech. Its Music 2.0 and 2.5 models supported compositions up to five minutes long.[5]

MiniMax launched its Agent workspace in January 2026, productizing capabilities it had used internally for software development, data analysis, operations, recruitment and sales and marketing. M2.5 followed in February, targeting coding, tool use and workplace applications. In March, CEO Yan Junjie announced an intention to evolve MiniMax “from a large-model company into a platform company for the AI era.”[5]

Released on March 18, 2026, M2.7 brought MiniMax’s agent work into its own research workflow. Researchers discussed experimental ideas with an agent and set experiment specifications; the agent assisted with literature review, prepared data and artifacts, launched experiments, monitored progress, analyzed metrics, debugged problems and proposed code fixes, merge requests and smoke tests. MiniMax said M2.7 could handle 30%–50% of this RL-team workflow, with humans retaining critical decisions and discussions.[9]

In a separate internal experiment, MiniMax had M2.7 optimize a programming scaffold over more than 100 autonomous rounds: analyze failures, plan changes, edit scaffold code, evaluate results and keep or revert changes. The model searched sampling parameters, added guidance to look for similar bugs elsewhere after a fix, and introduced loop detection. MiniMax reported a 30% improvement on internal evaluation sets. This example changed the model’s surrounding execution system; it did not establish autonomous training of model weights.[9]

M2.7 also supported MiniMax’s exploration of interactive entertainment. The company open-sourced OpenRoom, a preliminary agent-harness demo that placed conversations in an interactive Web GUI, with visual feedback, scene interactions and characters engaging with their environment.[9]

MiniMax released M3 during the first six months of 2026, extending its coding, agentic-workflow and professional-work offerings. H3 followed shortly after that reporting period with open weights for video generation.[6]

A January 11, 2026 exchange filing announced full exercise of MiniMax’s IPO over-allotment option. Its subsequent financial releases identified the listed company as MiniMax Group Inc., HKEX: 00100.[11][5]

For 2025, MiniMax reported revenue of US$79.0 million, up 158.9% year over year, with more than 70% from international markets. AI-native products contributed US$53.1 million; Open Platform and other AI-based enterprise services contributed US$26.0 million. Adjusted net loss was US$250.9 million.[5]

The unaudited first-half 2026 results showed a shift toward platform services. Revenue reached US$116.6 million, exceeding the preceding full-year total. Open Platform and other AI-based enterprise services generated US$73.9 million, or 63.4% of revenue, compared with 30.3% in first-half 2025. AI-native products generated US$42.6 million. Adjusted net loss increased to US$293.0 million from US$138.7 million in the corresponding half-year.[6]

Both releases defined adjusted net loss as a non-IFRS measure adding back share-based payment expenses, fair value loss on financial liabilities and listing expenses.[5][6]

At December 31, 2025, MiniMax reported cumulatively serving more than 236 million users and 214,000 enterprise customers and developers. Its August 26, 2026 announcement reported more than 300 million users across over 200 countries and regions, and more than one million enterprises and developers across over 100 countries. These figures did not specify active or paying users. Yan separately reported that July 2026 token consumption had reached 20 times its January level.[5][6]

On February 23, 2026, Anthropic alleged that MiniMax had conducted a Claude distillation campaign involving more than 13 million exchanges, targeting agentic coding, tool use and orchestration. Anthropic said it attributed the campaign through request metadata and infrastructure indicators, and that MiniMax redirected nearly half its traffic within 24 hours of a new Anthropic model release. These were Anthropic’s allegations, not adjudicated findings; its account did not name the MiniMax model allegedly being trained.[10]