en.Wedoany.com Reported - Xiyu Technology's MiniMax plans to launch a new generation large model, internally codenamed M3 Pro, with a parameter scale of 2.7 trillion. The model is expected to be released and open-sourced as early as the third quarter of this year, focusing on complex reasoning, multi-step task processing, agent execution, and code generation.
The parameter scale of M3 Pro is significantly larger than that of MiniMax M3. MiniMax M3 has a total of 428B parameters, with 23B activated parameters, adopting a sparse activation structure that only calls upon a portion of parameters for computation during inference. If M3 Pro reaches the 2.7 trillion parameter level, its focus will not merely be on "making the model larger," but on enhancing capabilities in long-chain problem decomposition, tool invocation, complex reasoning paths, code engineering generation, multi-round execution, and task state maintenance through larger expert capacity, stronger task routing, and more complex reasoning pathways. For agent models, parameter scale must serve execution capability; the model needs to break down user requirements into steps, select tools, read context, generate intermediate results, and continue to revise the task path based on feedback.
MiniMax M3 was officially open-sourced on June 15, with a total of 428 billion parameters and 23 billion activated parameters. This model is designed for coding and agent tasks, featuring autonomous task decomposition, tool invocation, multi-step reasoning, and long-context processing capabilities, with code generation goals shifting from "being able to run" to "being deliverable."
The technical direction of M3 Pro will continue to revolve around agent tasks. Complex reasoning is not just about math problems or Q&A; it involves stringing together code repository understanding, requirement analysis, error localization, file modification, test execution, performance optimization, and result delivery into a continuous process. Multi-step tasks are not just about calling the model once to generate an answer, but having the model cycle through searching, reading, planning, invoking tools, executing operations, and checking results. If MiniMax open-sources M3 Pro, developers can test the execution capabilities of larger parameter models in local deployments, private cloud deployments, enterprise internal agent platforms, and code workflows.
Founded in early 2022, Xiyu Technology develops multimodal models and AI-native applications, serving over 300 million individual users across more than 200 countries and regions, and providing services to over 1 million enterprise clients and developers in more than 100 countries and regions. Its product portfolio includes multimodal large models, AI applications, open platforms, and developer tools, with model capabilities extending from text, images, video, and speech to code and agent tasks.
If M3 Pro is open-sourced on schedule, it will further expand model capacity and agent capabilities based on MiniMax M3. The technical chain will focus on the 2.7 trillion parameter model architecture, sparse activation strategy, long-context window, tool invocation interface, code task evaluation, multi-step reasoning stability, and open-source deployment costs; these factors will determine whether it can evolve from a "larger model" into a "stronger agent execution model."
