MiniMax announced M3, a new model it describes as combining coding and agentic capabilities, up to a 1M-token context window, and native image and video multimodality. The company says the model is available through its API now and that weights and a technical report will be open-sourced soon, but they were not yet...
MiniMax announced M3 as a new AI model that it says combines coding and agentic capabilities, long-context processing, and native multimodal input.
In a company blog post titled “MiniMax M3: Frontier Coding, 1M Context, Native Multimodality — All in One Model,” MiniMax says M3 supports up to a 1M-token context window and can process text alongside native image and video inputs.
MiniMax’s M3 product page adds an important qualification: the model supports “up to” a 1M-token context window, with a guaranteed minimum of 512K tokens. That distinction matters because maximum context length and consistently available context length can differ in real deployments.
The company frames M3 as a model for coding and agentic use cases. According to MiniMax’s blog post, the model is available through its API, and the company says the technical report and model weights will be open-sourced over the next 10 days.
TheQuery’s June 2 report describes MiniMax M3 as an open-weight model, while also noting that the weights had not yet been released. MiniMax’s own materials are consistent with that timing: the blog post says the technical report and weights will be open-sourced over the next 10 days, and the product page says M3 will soon be fully open-sourced on Hugging Face and GitHub.
That means M3 is best described, for now, as a model with announced open-weight plans rather than a model whose weights are already available for independent download and testing.
TheQuery’s report says MiniMax made benchmark claims involving SWE-bench Pro and included pricing comparisons against Claude Opus 4.7. Those are vendor and media-reported comparisons, and they should be treated as claims rather than independent validation unless reproduced by third parties.
MiniMax’s own blog post presents M3 as competitive in coding and agentic tasks, but the cited sources do not provide enough independent evidence to confirm broader market positioning. The most solid facts from the available sources are the announced feature set, API availability, stated context-window limits, multimodal support, and the company’s stated plan to release weights and a technical report.
If MiniMax follows through on releasing M3’s weights, the model could be relevant for developers and organizations that want a long-context, multimodal model outside a purely closed API setting. A 512K guaranteed context window and up to 1M tokens, if usable in practice, would make M3 suitable for tasks involving large codebases, long documents, or multi-file analysis.
The release also reflects a broader trend: model providers are increasingly packaging coding, agentic workflows, long context, and multimodality into a single product rather than treating them as separate capabilities.
For now, the key caveat is availability. MiniMax says M3 is accessible through its API, but the open-weight release remains pending according to MiniMax’s blog and product page, as well as TheQuery’s June 2 coverage.
MiniMax announced M3 as a new AI model that it says combines coding and agentic capabilities, long context processing, and native multimodal input.
MiniMax’s M3 product page adds an important qualification: the model supports “up to” a 1M token context window, with a guaranteed minimum of 512K tokens.
That distinction matters because maximum context length and consistently available context length can differ in real deployments.
Continue reading
