MiniMax M3: New Model with 1M Tokens, Multimodality, and Computer Use

On June 1, 2026, MiniMax released a new model, MiniMax M3, built on the MiniMax Sparse Attention (MSA) architecture — with a context window of 1 million tokens and native support for images, video, and computer use mode.

How does the MSA architecture work?

Sparse Attention is an approach where a model does not recalculate attention across all pairs of tokens in a sequence, as in classical dense attention, but concentrates on a sparse subset of the most relevant connections. This makes it possible to process significantly longer sequences within the same computational budget. It is on this principle that MiniMax Sparse Attention is built — an architecture that became a key engineering solution allowing MiniMax M3 to maintain a context window of 1 million tokens without exponential growth in inference costs.

In a classical dense transformer, the computational complexity of the attention mechanism grows quadratically with the length of the sequence: doubling the context means a fourfold increase in required computations. This quadratic growth has historically been the main technical barrier to models with million-token context windows — without sparse or other optimized attention schemes, inference at such lengths becomes economically impractical. Architectural solutions like MSA allow labs to bypass this limitation by reducing the share of computations devoted to rarely used connections between distant tokens.

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

Multimodality and computer use "out of the box"

MiniMax M3 is announced as a model with native, not overlay, support for images and video — these modalities are processed by the model directly, not through separate adapters or conversion to text descriptions. Additionally, the model supports computer use mode — the ability to perceive the state of a computer screen and perform actions (clicks, text input, interface navigation) for autonomous task execution in graphical applications. Combined with a long context, this opens up scenarios where the model can hold large volumes of visual and textual context while executing multi-step agent tasks.

Why does a model need a million tokens of context?

A context window of 1 million tokens is comparable to the size of a large codebase, extensive documentation, or many hours of video recording in text description. For agentic coding tasks — the second ability announced in the release — long context is critical: an agent working on a real project often needs to simultaneously hold in memory many files, change history, and results of command execution without losing the thread of the task.

MiniMax is a Chinese company developing large language models and competing in the market with other major developers from China and the United States, for whom long context and multimodality in recent years have become one of the main fields of competition alongside pure answer quality. The support for agentic coding announced in the release indicates that MiniMax M3 is positioned not simply as a dialogue model but as a tool to be embedded in agent development pipelines — that is, in the same class of tasks where models underlying products like autonomous coding assistants currently compete.

The release of MiniMax M3 fits into the general race of major labs for long context and universal multimodality — a direction in which, besides MiniMax, other model developers actively compete with context windows from hundreds of thousands to over a million tokens. The bet on sparse attention as a way to cheapen the processing of long sequences is one of the architectural answers that different labs solve in their own way.

For end users and developers, the practical significance of such releases is measured not so much by the announced architectural details as by what scenarios become truly available: processing an entire code repository in one pass, analyzing long video without dividing into fragments, or an agent that can keep in context the entire history of a multi-hour work session on a task. It is precisely the combination of long context with native multimodality and computer use in a single model, rather than in a set of disparate tools, that distinguishes releases like MiniMax M3 from earlier generations of models where each of these capabilities had to be assembled from separate components.