MiniMax M3: a new model with 1M-token context, multimodality and computer use

On June 1, 2026, MiniMax released a new model, MiniMax M3, built on the MiniMax Sparse Attention (MSA) architecture. It has a 1 million-token context window and native support for images, video and computer use.

How does the MSA architecture work?

Sparse attention is an approach in which the model does not recompute attention between every pair of tokens in a sequence, as classic dense attention does. Instead, it focuses on a sparse subset of the most relevant connections. This makes it possible to process much longer sequences on the same compute budget. MiniMax Sparse Attention is built on this principle, and it is the key engineering decision that let MiniMax M3 hold a 1 million-token context window without an exponential rise in inference costs.

In a classic dense transformer, the computational complexity of attention grows quadratically with sequence length: doubling the context means four times the required computation. This quadratic growth has historically been the main technical barrier to models with million-token context windows. Without sparse or otherwise optimized attention schemes, inference at such lengths becomes economically impractical. Architectures like MSA let labs work around this limit by cutting the share of computation spent on rarely used connections between tokens that are far apart.

• Model: MiniMax M3
• Architecture: MiniMax Sparse Attention (MSA)
• Context window: 1 million tokens
• Modalities: native images, video, computer use
• Also: agentic coding support
• Release date: June 1, 2026 (according to MarkTechPost)

Multimodality and computer use out of the box

MiniMax M3 is billed as having native rather than bolted-on support for images and video. The model processes these modalities directly, not through separate adapters or by converting them into text descriptions. The model also supports computer use: it can perceive the state of a computer screen and take actions (clicks, text input, navigating interfaces) to complete tasks autonomously in graphical applications. Combined with a long context, this opens up scenarios where the model can keep large amounts of visual and textual context in memory while carrying out multistep agentic tasks.

Why does a model need a million tokens of context?

A 1 million-token context window is comparable to the size of a large codebase, extensive documentation, or a multi-hour video recording rendered as a text description. For agentic coding, the second capability the release claims for the model, a long context is critical: an agent working on a real project often has to hold many files, the change history and command outputs in memory at once without losing the thread of the task.

MiniMax is a Chinese company that develops large language models and competes with other major developers in China and the US. For all of them, long context and multimodality have become one of the main battlegrounds in recent years, alongside raw answer quality. The agentic coding support claimed in the release suggests that MiniMax M3 is positioned not merely as a conversational model but as a tool to embed in agentic development pipelines, the same class of tasks where the models behind autonomous coding assistants now compete.

The release of MiniMax M3 fits into a broader race among large labs for long context and general-purpose multimodality. Besides MiniMax, other model developers are actively competing with context windows ranging from several hundred thousand to more than a million tokens. Betting on sparse attention as a way to make long sequences cheaper to process is one of the architectural answers that different labs approach in their own ways.

For end users and developers, the practical value of such releases is measured less by the claimed architectural details than by which scenarios become genuinely available: processing an entire code repository in one pass, analyzing a long video without cutting it into fragments, or an agent that can keep the whole history of a multi-hour working session in context. What sets releases like MiniMax M3 apart from earlier generations of models, where each of these capabilities had to be assembled from separate components, is the combination of long context, native multimodality and computer use in a single model rather than in a set of disparate tools.