MiniMax's new M3 AI model can outperform OpenAI GPT-5.5
What's the story
Chinese artificial intelligence start-up, MiniMax, has launched its flagship AI model, M3. The Shanghai-based company introduced the new model today as part of its expansion into coding agents and automated workflows. The innovative architecture of M3 drastically reduces computational requirements by up to 95%, significantly cutting down inference costs while improving response speeds.
Advanced features
M3 can process 1M tokens at once
The new model, M3, can handle up to one million tokens of data at a time, five times more than its predecessor M2.7. This improvement enables the model to handle long and complex programming tasks. In one such test, MiniMax said that M3 was able to optimize software running on NVIDIA's Hopper chips.
Performance comparison
What about its performance?
In a major coding benchmark, SWE-Bench Pro, MiniMax-M3 outperformed rival models such as OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro. This shows that the new model is highly capable of handling software engineering and automated tasks. However, the company has not revealed the size of M3 or the computing infrastructure it used for training purposes.
Strategic move
MiniMax's IPO plans
The launch of MiniMax M3 is the company's first major product release since it officially started preparing for an initial public offering (IPO) on Shanghai's tech-heavy Star Market. The move will complement its existing listing in Hong Kong.