MiniMax M3 Preview

• 🧠 Frontier 1.4-trillion-parameter MiniMax model for coding, agents, reasoning.
• 📏 512K-token context window in this preview, served at fp8.
• 🆕 Built on new MiniMax Sparse Attention (MSA) for long context.
• ⚡ MSA enables efficient native ultra-long-context pretraining.
• 🔧 Function calling, tool use, and structured agentic task execution.
• 👁️ Sibling M3 is natively multimodal (text, image, video input).
• 🌐 Web search and long-horizon agentic workflows supported.
• 🎯 Targets autonomous coding and multi-step agentic reasoning.

MiniMax

MiniMax is an AI company building generative models across multiple modalities, with a focus that spans both language understanding and audio creation. Their rapid release cadence in early 2026—delivering several new models within just a few months—reflects…

MiniMax M3 Preview is a preview build of MiniMax's flagship M-series language model, described in this catalog as a 1.4-trillion-parameter frontier model for coding, agentic workflows, and complex reasoning, served at fp8 with a 512K-token context window. It is positioned alongside the full MiniMax M3 release, which MiniMax presents as a model combining frontier coding, ultra-long context, and native multimodal input in a single architecture.

The central change from earlier M-series models is MSA (MiniMax Sparse Attention), which replaces the quadratic cost of full attention to enable native ultra-long-context pretraining, according to MiniMax. The production M3 supports up to 1M tokens with a guaranteed 512K minimum; this preview exposes the 512K tier.

Compared with prior family members such as MiniMax M2.7 and MiniMax M2.5, which remain available for existing workflows, MiniMax frames coding and agentic capability as M3's key areas of improvement, with autonomous task decomposition, tool invocation, and multi-step reasoning.

As a function-calling and web-search-capable model, M3 Preview is aimed at long-horizon agentic and computer-use tasks, including code generation and tool-driven workflows. Being a preview, weights, the technical report, and full availability above 512K tokens were still being rolled out around launch, with the model exposed here at the 512K context tier.