HeadsUpAI

Arena adds MiniMax M3 for community testing of 1M context model

· Updated

Arena added MiniMax M3 to its community evaluation platform, marking the arrival of the first open-weights model to combine frontier coding with native multimodality. Developed by MiniMax, the model features a 1-million-token context window and is now live for testing across text, vision, and frontend coding categories.
Model
MiniMax M3
Context Window
1,000,000 tokens
Arena Categories
Text, Vision, Document, and Code
SWE-Bench Pro
59.0%
MCP Atlas
74.2%

The model matches closed-source leaders in software development, scoring 59.0% on SWE-Bench Pro. It builds on the agentic capabilities of MiniMax M2.7 to provide a more powerful engine for autonomous tasks. This context window allows agents to navigate entire repositories, mirroring the scale seen in DeepSeek-V4.

Access is available now via the MiniMax API and the MiniMax Code platform, with standard usage under 512K tokens discounted by 50% for the first week. Users can also test prompts in the Arena to vote on performance, with official scores expected soon, before the weights and technical report arrive in roughly 10 days.

View the full update on platform.minimax.io
Arena.ai
Arena.ai
@arena
X

New open model: MiniMax M3 by @MiniMax_AI is live in the Arena! Find it across Text, Vision, Document and Code Arena: Frontend. Bring your toughest prompts and vote. Scores incoming soon! https://t.co/aF2kpvLyWi

16retweets262likes
View on X

Still wondering? A few quick answers below.

MiniMax M3 is a multimodal foundation model designed for high-performance coding and autonomous agentic tasks. It is the first open-weights model to combine a 1-million-token context window with native support for text, image, and video inputs, achieving frontier-level scores on software engineering benchmarks like SWE-Bench Pro.

The model utilizes MiniMax Sparse Attention, a specialized architecture designed to scale context efficiently. By focusing on the most relevant data points within a massive input rather than processing every token equally, the system can maintain performance and accuracy across 1 million tokens without the computational overhead of standard attention mechanisms.

While the model is currently accessible via the MiniMax API and the MiniMax Code platform, the company has announced that the open weights and a detailed technical report will be released in approximately 10 days. This will allow developers to run and fine-tune the model on their own infrastructure.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Share this update