MiniMax-M2.1

MiniMax-M2.1 will be released in December 2025, using a hybrid expert architecture (MoE), with a total parameter volume of approximately 230B and activation parameters of approximately 10B. Its core positioning is to break the shortcomings of mainstream AI programming tools in multi-language adaptation and complex task execution. It systematically strengthens the programming capabilities of Rust, Java, Golang, C++, Kotlin, TypeScript and other languages, covering the complete link from back-end systems to mobile native development. M2.1 introduces the ability to perceive design aesthetics and can generate interactive interfaces and 3D scenes that are both functional and visually expressive, making "vibe coding" a deliverable production practice. Compared with M2, M2.1's reply and thinking chain are simpler, the response speed is significantly improved, and the Token consumption is significantly reduced. In the agent scenario, M2.1 shows stable generalization capabilities under the framework of mainstream programming tools such as Claude Code and Cline.

Context

204800 tokens

Max output

32768 tokens

Input

llm

Output

llm