MiniMax M3 combines agentic code, 1M context, and native multimodality

MiniMax M3 debuts as an open-weights model with a 1M context window, native multimodality, and 59% on SWE-Bench Pro, beating GPT-5.5 and Gemini 3.1 Pro.

MiniMax AI introduces M3, its new flagship model. The company positions it as the first open-weights model to combine in a single system three so-called frontier capabilities: performance on code and autonomous agent tasks, a one-million-token context window, and native multimodality acquired during training.

In coding and agentic tasks, M3 claims 59% on the SWE-Bench Pro benchmark, ahead of GPT-5.5 and Gemini 3.1 Pro and neck-and-neck with Claude Opus 4.7, as well as strong results on several agent and tool-calling evaluations. The extended context relies on an architecture named MiniMax Sparse Attention, which replaces classic dense attention and divides the computational cost per token by twenty at one million tokens, with pre-filling and decoding also significantly accelerated. As input, the model accepts text, images, and videos (up to 1,024 images) and can control a computer via the in-house tool MiniMax Code.

The API is immediately accessible at platform.minimax.io, at a price point between 5% and 20% of comparable US models, and already supported by several gateways such as Vercel, Cloudflare, or OpenRouter. MiniMax Code, a desktop application that leverages M3, adds persistent agents and agent teams. The release of the weights and a technical report is announced for about ten days, with the model then becoming open, following the example of the M2 versions that preceded it.