MiniMax just released M3, positioning it squarely against the top coding models with a 1 million token context window and native support for agentic workflows.
A 1 million token context window means the model can hold roughly 750,000 words in a single session - enough to ingest an entire large codebase, multiple documentation files, and a long conversation history all at once without losing track of earlier details. That scale matters most for developers working on multi-file projects or for AI agents (autonomous programs that plan and execute tasks in steps rather than just answering a single question) that need to read, reason about, and modify code across many files in one pass.
The multimodal angle - meaning M3 can process both text and images, not just text - extends its usefulness beyond pure code generation. Feeding screenshots of UI mockups, error messages, or architecture diagrams directly into the model is the obvious use case for developers and product teams.
MiniMax is a Chinese AI lab that has been quietly shipping competitive models. M3 enters a crowded coding-model bracket that currently includes Claudee Code](/tools/claude-code/), Cursor's underlying stack, and Google's Gemini 2.5 Pro, all of which also compete on long context. The 1M token figure matches Gemini 2.5 Pro's headline number, though real-world performance at the edges of a long context window varies significantly between models and is only reliably tested with structured benchmarks.
No pricing details or API availability timeline have been confirmed publicly at launch. Given MiniMax's track record of releasing models on Hugging Face for local use as well as via API, weights availability is likely to follow.