MiniMax has announced MiniMax M3, a model it describes as combining coding, agentic capabilities, native multimodality, and a 1M-token context window. The company says M3 is available through its API, while OpenRouter lists it as supporting text, image, and video inputs.
MiniMax announced MiniMax M3 on June 1, 2026, describing the model as a new release aimed at coding, agentic workflows, long-context tasks, and multimodal use cases.
In its launch post, MiniMax says M3 combines “frontier coding,” a 1M-token context window, and native multimodality in one model. The company’s M3 product page similarly describes the model as an open-weight system built for coding and agentic capabilities, with a 1M-context MSA architecture and multimodal support.
MiniMax’s positioning places M3 in a competitive category of general-purpose models that are expected to work across software engineering, document-heavy workflows, and multimodal inputs. The company specifically points to benchmark claims across SWE-Bench Pro and agentic tasks in its launch materials, although independent evaluations will be important for assessing how those claims translate into real-world development environments.
The headline technical feature in MiniMax’s announcement is the 1M-token context window. A context window of that size is intended to let a model process very large inputs, such as long codebases, extensive documentation, multi-file project histories, or lengthy mixed-media task instructions.
MiniMax also provides tiered token-plan and API pricing details for long-context use, according to its launch post. That matters because very large context windows can be expensive to use in practice, particularly for workloads that repeatedly send long prompts or large project states to a model.
OpenRouter’s listing for minimax/minimax-m3 also describes the model as having a 1M-token context window and an MSA architecture. OpenRouter lists the model with a May 31, 2026 release date, while MiniMax’s own launch post says M3 was released on June 1, 2026. The one-day difference may reflect listing timing, time zones, or publication timing across platforms.
MiniMax describes M3 as natively multimodal. Its product page says the model combines coding and agentic capabilities with native multimodality, while OpenRouter’s model page lists support for text, image, and video inputs.
That combination could make M3 relevant for workflows that combine code, screenshots, design assets, diagrams, and video-based context. For example, multimodal capability may be useful in debugging visual interfaces, reviewing product behavior from recordings, or connecting documentation with implementation details. The provided sources do not specify output modalities, so it is safest to describe the model’s multimodality in terms of the listed inputs rather than assuming image or video generation support.
MiniMax says M3 is available through its API. The company’s launch post also includes pricing details for token plans and API usage, particularly relevant for long-context workloads. OpenRouter lists the model as available under the minimax/minimax-m3 identifier and provides its own model-page details, including pricing and input support.
The MiniMax product page describes M3 as an open-weight model. That label is significant for developers and organizations that want more control over deployment or inspection than they typically get from closed hosted-only models. However, the provided source excerpts do not include the full license terms, download details, or deployment requirements, so developers should review MiniMax’s official model page before assuming how the weights can be used commercially or redistributed.
M3’s announcement reflects continuing competition around models that combine long context, coding performance, tool-using or agentic behavior, and multimodal inputs. MiniMax is presenting the model as a single system for several high-demand use cases rather than as a narrow coding model or a standalone multimodal assistant.
For teams evaluating M3, the key questions will be practical ones: how well it performs on their own repositories, how reliably it handles long-context retrieval and reasoning, how costly 1M-token use becomes at scale, and whether its multimodal input support fits their workflows. MiniMax’s benchmark and product claims provide a starting point, but production testing and third-party comparisons will be needed before drawing broader conclusions about where M3 stands against other frontier models.
MiniMax announced MiniMax M3 on June 1, 2026, describing the model as a new release aimed at coding, agentic workflows, long context tasks, and multimodal use cases.
What MiniMax says M3 is designed to do In its launch post, MiniMax says M3 combines “frontier coding,” a 1M token context window, and native multimodality in one model.
The company’s M3 product page similarly describes the model as an open weight system built for coding and agentic capabilities, with a 1M context MSA architecture and multimodal support.
Continue reading
