Copyright scmp
MiniMax, a leading Chinese artificial intelligence start-up, released on Monday the M2, which ranks as the world’s leading open model on a prominent leaderboard. MiniMax M2 achieved an unprecedented score for an open model on Artificial Analysis’s overall intelligence index, placing it among the top five models globally. It surpassed Google DeepMind’s Gemini 2.5 Pro, released in June, and sat below leading models from US companies OpenAI, Anthropic and xAI. With 200 billion parameters, MiniMax M2 adopts the Mixture-of-Experts architecture, which has significantly enhanced the efficiency of Chinese AI models over the past year. The model activates only 10 billion of those parameters for each “forward pass” – the process of input data handling – continuing a trend among Chinese developers to improve efficiency. Using only a small subset of parameters allowed the model to operate efficiently at scale, according to Artificial Analysis. For comparison, DeepSeek’s V3.2 model and Moonshot AI’s Kimi K2 use 37 billion and 32 billion active parameters, respectively. “I am really impressed by their progress,” said Florian Brand, a PhD student at Germany’s Trier University and an expert on open models, highlighting substantial improvements from the previous M1 model. M2 excels in reasoning through complex agentic tasks, making it particularly adept at agentic and coding applications, according to Skyler Miao, head of engineering at MiniMax. “Our model is designed for thinking,” Miao said. “In multi-turn agentic tasks, it needs to carry over the thinking content from each response into the next request to maintain a consistent reasoning chain.” Other Chinese AI start-ups, including DeepSeek and Moonshot AI, have also prioritised developing agentic capabilities in recent months, aiming to build AI systems capable of autonomously performing complex tasks for users. The M2 is available through MiniMax’s application programming interface, or API, at a rate of US$0.30 per million input tokens and US$1.20 per million output tokens, considerably lower than its US rivals. However, the model consumed a substantial 120 million tokens to complete Artificial Analysis’ benchmark evaluations, making it one of the most token-intensive models, alongside xAI’s Grok 4. MiniMax’s previous M1 model was similarly token-heavy, requiring “many tokens for the simplest problems”, Brand said. Shanghai-based MiniMax is among Chinese AI start-ups that are aggressively pursuing international users, bolstered by its popular AI companion app, Talkie, and video generation app, Hailuo AI. On Monday, global model inference provider vLLM said it would offer support for the M2 immediately on the launch date. MiniMax is eyeing a potential initial public offering in Hong Kong as soon as this year, with a target valuation of more than US$4 billion, according to a Reuters report in July.
