Chinese AI startup MiniMax, once a pioneer in large language models, is now leaning on price rather than raw performance to compete as it pivots its business toward enterprise customers, according to a report by TMTPost.
The company, founded in late 2021 and listed in Hong Kong in January 2026, saw its market value briefly exceed HK$410 billion (about $57.7 billion) before a sharp pullback. Its M3 language model ranks 17th globally on the Artificial Analysis intelligence index, behind overseas models like Claude Fable 5.1 and GPT-6 Astra, and even behind Chinese rivals such as Zhipu GLM-5.3, Kimi K3, Alibaba Qwen 3.8, and DeepSeek V4 Pro.
Full-modality bet, price-led strategy
MiniMax is unusual among startups for investing across text, voice, video, and music models. Its H3 video model is particularly strong, ranking first for image-to-video on Artificial Analysis, and second for video editing, third for text-to-video. The H3 model integrates multiple tasks into a single framework, allowing users to mix up to 12 reference materials in one prompt, and it handles text rendering in videos better than previous generations.
However, the H3 model competes mainly on cost. At 1080p, rivals like Seedance 2.5 and Veo 3.1 charge about 3 yuan per second, while Wan 3.0, Kling 3.0, and Gemini Omni Flash charge about 1.1 yuan. MiniMax H3, at 2K resolution, costs just 0.8 yuan per second. On Reddit, users describe it as fast, cheap, and the most reliable in its price range. A short-drama producer told TMTPost that while Seedance gives the best results, they use MiniMax H3 more often because it delivers 80-90% of the quality at a much lower price.
Revenue shift to B2B
MiniMax's revenue mix has changed dramatically. In 2024, consumer AI products accounted for 71.4% of revenue, falling to 67.2% in 2025, as total revenue grew from $30.5 million to $79.0 million. But in the first half of 2026, consumer products dropped to 36.6% of revenue, while open platform and enterprise services surged 700% year-on-year to $73.9 million, making up 63.4% of the $116 million in total revenue.
The shift reflects the difficulty of monetizing consumer apps like Talkie, a social product where users create AI characters, and Hailuo AI, which relies on viral hits. Enterprise API sales, by contrast, directly tie revenue to model capability and pricing. MiniMax's M3 costs just $0.51 per task on Artificial Analysis, cheaper than DeepSeek's $0.67 and far below OpenAI's $3.26 or Anthropic's $8.75.
CEO Yan Junjie said on an August earnings call that annual recurring revenue exceeded $800 million, and that price cuts and margin improvement are not contradictory if per-token gross margin stays positive. Yet the company remains loss-making, squeezed by low prices and high computing costs for full-modality inference.
