MiniMax has introduced MiniMax M3, an open-weight model described by the company as combining coding, agentic reasoning, a 1 million-token context window, and native multimodal capabilities.
MiniMax introduced MiniMax M3, an open-weight model that the company says is designed for coding, agentic reasoning, long-context work, and multimodal tasks.
In its announcement, MiniMax describes M3 as a single model that combines “frontier coding” and agentic capabilities with a 1 million-token context window and native multimodality. The company’s official model page similarly presents M3 as an open-weight model built around coding, agentic tasks, a 1M-context MSA architecture, and multimodal input support.
MiniMax’s API documentation also lists MiniMax-M3 as available through the company’s platform with a 1,000,000-token context window. According to the API overview, the model is intended for agentic reasoning, tool use, coding, and long-context tasks.
The main emphasis in MiniMax’s materials is on software-development and agent workflows. MiniMax says M3 is built for coding and agentic capabilities, and its API documentation specifically calls out tool use and long-context tasks as supported use cases.
MiniMax also cites benchmark performance in its announcement, including a reported 59.0% score on SWE-Bench Pro. That figure comes from MiniMax’s own blog post, so it should be read as a vendor-reported benchmark result unless independently reproduced.
MiniMax says M3 supports a 1 million-token context window. The company’s model page describes this as part of a “1M-context MSA architecture,” while the API documentation lists the available context length as 1,000,000 tokens.
For developers, the practical significance of such a long context window is that the model may be able to process much larger inputs in a single request than shorter-context models. MiniMax’s own documentation points to long-context tasks, agentic reasoning, tool use, and coding as target applications.
MiniMax describes M3 as having native multimodality. The official model page uses the same framing, saying the model combines coding and agentic capabilities with native multimodal support. The provided MiniMax sources do not specify all supported input or output modalities in the excerpts, so the exact scope of multimodal functionality should be checked in the full model and API documentation.
MiniMax also says M3 is an open-weight model. The company’s model page uses that term directly, indicating that the model is being positioned for broader developer access than a closed-only hosted model. The API documentation separately lists MiniMax-M3 as available on MiniMax’s platform.
MiniMax’s announcement says M3 uses length-dependent pricing above 512K input tokens. The provided sources do not include the full pricing table, but the note indicates that very long requests may be priced differently once input length passes that threshold.
M3 is notable because MiniMax is presenting several capabilities in one model: coding, agentic reasoning, tool use, long-context processing, and native multimodality. The company’s own sources consistently describe a 1 million-token context window and availability through MiniMax’s API, while the blog post adds vendor-reported benchmark and pricing details.
As with any model launch, the most important follow-up questions are how the model performs in independent evaluations, how its long-context behavior holds up in real applications, and how pricing affects workloads that approach the upper end of the 1 million-token context window.
MiniMax introduced MiniMax M3, an open weight model that the company says is designed for coding, agentic reasoning, long context work, and multimodal tasks.
What MiniMax is claiming In its announcement, MiniMax describes M3 as a single model that combines “frontier coding” and agentic capabilities with a 1 million token context window and native multimodality.
The company’s official model page similarly presents M3 as an open weight model built around coding, agentic tasks, a 1M context MSA architecture, and multimodal input support.
Continue reading
