MiniMax releases M3 with 1M context and native multimodal input

MiniMax has released M3 as an open-weight model for coding agents, long-context work, and native multimodal understanding. The official model card lists roughly 428B total parameters, about 23B active parameters, and input support across text, images, and video.

The launch matters because M3 is not only another model-directory update. MiniMax is positioning it as the first open-weight model that combines three capabilities usually associated with closed frontier systems: million-token context, agentic coding strength, and native multimodal input.

MiniMax M3 targets long-running agent work

MiniMax says M3 was trained for coding and agentic tasks where a model has to keep state across long sessions, inspect files, use tools, and recover from intermediate failures. Its product page highlights autonomous task decomposition, tool invocation, and multi-step reasoning as core use cases rather than add-ons.

The headline capability is context. MiniMax says the M3 API supports up to a 1M-token context window, with a guaranteed minimum of 512K tokens. OpenRouter's public model listing also shows a 1,048,576-token context length for minimax/minimax-m3, while its listed top provider currently exposes a 524,288-token context. That difference is worth checking before production routing, because provider limits can be lower than the model's theoretical or first-party ceiling.

Sparse attention is the technical bet

The architecture change behind M3 is MiniMax Sparse Attention, or MSA. MiniMax describes MSA as a sparse attention design built to reduce the cost of million-token contexts without discarding too much relevant history.

In the announcement, MiniMax says MSA reduces M3's per-token compute at a 1M context length to 1/20 of its previous-generation model, with more than 9x faster prefill and more than 15x faster decoding in its internal comparisons. Those are vendor-reported numbers, so readers should treat them as launch claims until independent serving benchmarks arrive.

The model card adds the deployment-facing details: M3 is published as safetensors, supports Transformers, vLLM, SGLang, Docker Model Runner, and other local or hosted paths, and exposes configurable reasoning modes. The Hugging Face page also shows live third-party inference providers, which means developers can test the model without waiting for a private enterprise channel.

The model is multimodal from the start

M3 is listed as an image-text-to-text model on Hugging Face, but MiniMax's own copy is broader: it says the model was trained with mixed-modality data from the beginning and supports text, image, and video input. OpenRouter lists the architecture as text+image+video->text.

That matters for coding and agent workflows. A model that can read screenshots, UI states, diagrams, documents, and video frames inside a long task thread can handle a wider class of software work than a pure text model. MiniMax uses examples such as paper reproduction, CUDA kernel optimization, and office-style automation to show where long context and visual understanding overlap.

The practical caveat is output modality. This is not an image or video generator. Public listings describe text output, so the useful lane is multimodal analysis and agent execution rather than media creation.

Access, pricing, and open-weight status

MiniMax says M3 is available through MiniMax Code, Token Plan subscriptions, and API services. Its launch post lists token-plan tiers and says text, image, speech, and music share the same usage pool. The product page also says M3 will be open-sourced on Hugging Face and GitHub for private deployment and fine-tuning.

The artifact side is already visible. Hugging Face lists the MiniMaxAI/MiniMax-M3 repository as public and not gated, with safetensors files and a license file. GitHub carries the matching MiniMax-AI/MiniMax-M3 repository with the same core model description.

For LinkLoot readers building agent stacks, this is a candidate to test against closed coding models on repository-scale tasks, screenshot-heavy debugging, and long document or video analysis. Start with a narrow benchmark that matches your workflow, then compare context reliability, tool-call behavior, latency, and provider-specific limits. Our broader agent tooling guide is here: AI agent tools.

Sources and verification notes

This post uses MiniMax's announcement and model page as primary sources, with Hugging Face, GitHub, and OpenRouter as public artifact and availability checks. Claims about benchmark leadership, long autonomous runs, and speedups are presented as MiniMax-reported unless they are directly observable from public model listings.

The strongest confirmed facts are the public model artifact, approximate parameter scale, native multimodal input, 1M-context positioning, and current third-party availability. The open question is how M3 holds up under independent, repeatable agent benchmarks once teams test it outside MiniMax's launch harness.