MiniMax M3 is officially released today. This model reaches frontier level on coding and Agent tasks, ships with the brand-new sparse attention architecture MSA (MiniMax Sparse Attention), supports up to 1M of ultra-long context, and is a natively multimodal model that supports image and video input and can operate the computer desktop.

M3 is the first model in China to combine all three of frontier Coding capability, ultra-long context, and native multimodality.

MiniMax M3's results across multiple benchmarks

Core Capabilities at a Glance

Coding and Agent Capability

On internationally authoritative evaluations of Coding capability, M3 reaches world-leading level:

• SWE-Bench Pro: 59.0% (surpassing GPT-5.5 and Gemini 3.1 Pro, close to Opus 4.7)
• Terminal Bench 2.1: 66.0%
• SWE-fficiency: 34.8%
• KernelBench Hard: 28.8%
• MCP Atlas: 74.2%
• SVG-Bench: surpasses Opus 4.7

M3 also scored highest on Claw-Eval, the Agent end-to-end evaluation framework.

The New Attention Architecture MSA

MSA is MiniMax's in-house sparse attention architecture, solving the problem of quadratic growth in computational complexity in full attention. Its core advantages:

• Precise KV chunking: More precise than DSA and MoBA, achieving higher effective context coverage
• Compute and memory-access optimization: Uses the KV outer gather Q approach, reading each block only once with contiguous memory access
• Significant speedup: More than 4x faster than the open-source Flash-Sparse-Attention and FlashMoBA

Speed comparison of MSA against other sparse attention solutions

At 1 million context, M3's per-token compute is only 1/20 of the previous-generation model. Prefilling is sped up by more than 9x, and Decoding by more than 15x.

Native Multimodality

M3 performs multimodal mixed training from Step 0 — instead of attaching a vision encoder after the fact, the semantic spaces of different modalities of data fuse naturally. Interleaved data in the training corpus proved especially critical for the performance gains, and the training token scale has been raised to the order of 100 trillion.

Hands-On: Reproducing an Award-Winning Paper in 12 Hours Unattended

The MiniMax team handed M3 an ICLR 2025 Outstanding Paper Award winner — Learning Dynamics of LLM Finetuning — and asked it to reproduce the work independently.

M3 ran autonomously for nearly 12 hours, unattended throughout, and ultimately:

• Produced 18 commits and 23 experiment charts
• Successfully matched the predicted probability trends in the SFT stage
• Clearly observed the squeezing effect in the DPO experiments
• Validated the Extend mitigation method proposed in the original paper

This test simultaneously drew on M3's three strengths: 1M ultra-long context (reading the full paper), top-tier coding capability (writing experiment code), and native multimodality (generating and interpreting experiment charts).

How to Try It

You can experience MiniMax M3 first-hand on these platforms:

• MiniMax Code: Online coding environment
• Token Plan: API calling service
• MiniMax API: Direct integration into your project

The Interactive User Simulator Framework

M3's Coding gains don't come from Benchmark training alone. MiniMax built an interactive user simulator framework that models how real developers behave during collaboration:

• Requirement clarification and solution discussion
• Feedback correction and continuous task switching
• Multi-round iterative optimization on complex projects

This lets the Agent stop passively executing instructions and instead actively collaborate with the user to get tasks done. The next generation of Agent Coding will compete not just on code generation, but on long-horizon collaboration and human-Agent coordination efficiency.

Who It's For

• Developers: Need an Agent with strong Coding capability to handle complex software engineering tasks
• Researchers: Need to process ultra-long documents (papers, codebases) and perform complex reasoning
• Agent builders: Need a foundation model that can simultaneously understand text, images, and video and operate a computer
• Enterprise users: Need to handle long-context enterprise knowledge bases and document analysis scenarios