MiniMax M3 is an open-weight Mixture-of-Experts model for frontier coding and agentic work. It runs 23B active parameters out of 428B, takes a 1M-token context, and is natively multimodal: text, image, and video in, text out. It's built for long autonomous runs that read a whole repository, plan, call tools, and keep going across hundreds of steps.
M3 is the first open-weight model to combine all three of these at once: frontier coding, a million-token context, and native multimodality. Until now that combination was closed-source only.
With is*smart, M3 is already deployed and optimized inside is*hosting, so you skip the hardware and setup. Here is where it earns its place.
This is the model's headline strength. M3 takes a large task, breaks it down, edits across files, runs terminal commands, and iterates over long sessions without drifting. It's the option for the hard, multi-hour jobs that shorter-running models abandon halfway.
The 1M-token window plus native image and video input changes what you can hand it. Give it the repository, the failing tests, a screenshot of the broken UI, or a screen recording of the bug, and it reasons over all of it together instead of one text file at a time.
M3 scores well on web-browsing and retrieval benchmarks, which makes it a solid base for agents that search, read across sources, and synthesize an answer. It holds context across many pages without losing the original question.
Because M3 can operate a desktop, it fits UI automation and tool-driven workflows: clicking through an interface, filling forms, moving data between apps, and driving software that has no clean API.
M3 is the heavy option in the is*ai lineup. When you don't need frontier capability, a lighter model is cheaper and faster:
M3 is a frontier model, and that capability costs more compute than a Flash-tier model, so it's overkill for routine or high-volume calls. Long autonomous runs also burn a lot of tokens, so scope the task and set limits before you turn it loose. The context window is guaranteed to 512K and stretches to 1M depending on setup, so keep prompts clean at extreme length. And note the output is text only: M3 reads images and video but does not generate them.
Open weights are only useful if you can run them. Through is*smart, M3 is already hosted and optimized inside the is*hosting infrastructure, so your code, prompts, and files stay in the environment, with no third-party APIs and no data leaving. You get frontier coding and agentic capability without sourcing GPUs or managing weights yourself.
Subscribe to is*smart to get instant access to MiniMax M3 and put it on your hardest, longest tasks.