What Is MiniMax M3?
MiniMax M3 is an open-weight AI model launched June 1, 2026 by MiniMax — a Shanghai-based AI lab founded in 2021. M3 is the first open-weight model to combine three frontier capabilities in one architecture: top-tier coding performance, a 1 million token context window, and native multimodal input (text, image, and video). It is built on MiniMax’s proprietary Sparse Attention (MSA) architecture, which enables it to process long contexts at approximately four times the speed of comparable open-source approaches.
MiniMax M3 Key Features
- 1M token context window: Process entire codebases, long documents, or video in a single pass — without chunking.
- Native multimodal: Text, image, and video inputs natively supported, including computer-use capabilities for agentic OS interactions.
- Frontier coding: 59.0% on SWE-Bench Pro — beating GPT-5.5 and Gemini 3.1 Pro on the leading coding benchmark.
- Open weights: Weights published on Hugging Face and GitHub — self-host on your own hardware, fine-tune, or audit the model.
- Agentic design: M3 underpins MiniMax’s multi-agent framework Mavis, supporting autonomous project execution and complex task orchestration.
- MSA architecture: Per-token compute cut to roughly 1/20th of the previous generation at equivalent context lengths.
MiniMax M3 Benchmarks
| Benchmark | MiniMax M3 | Notes |
|---|---|---|
| SWE-Bench Pro | 59.0% | Beats GPT-5.5 and Gemini 3.1 Pro |
| Terminal-Bench 2.1 | 66.0% | Strong coding agent result |
| SWE-fficiency | 34.8% | Efficient software engineering |
| KernelBench Hard | 28.8% | Low-level compute tasks |
| MCP Atlas | 74.2% | Agentic tool-use benchmark |
| BrowseComp | 83.5 | Beats Claude Opus 4.7 |
MiniMax M3 Pricing
| Access Method | Price |
|---|---|
| API (standard) | $0.60 / $2.40 per million in/out tokens |
| API (launch promo on OpenRouter) | ~$0.30 / $1.20 per million in/out tokens |
| Self-hosted (open weights) | Hardware cost only |
Who Should Use MiniMax M3?
M3 is an excellent choice for cost-conscious developers who need frontier coding performance without paying GPT-5.5 or Claude Opus rates. It is particularly strong for: agentic workflows requiring long context, document and video analysis tasks, teams who want to self-host for data privacy, and developers benchmarking open-weight alternatives to closed models. The key caveat: M3 is open-weight, not fully open-source — training code and inference components have not been released. Run independent validation on your specific tasks before committing to M3 in production.
Verdict
MiniMax M3 is the most significant open-weight model release of 2026 so far. It closes the gap between open and closed frontier AI more dramatically than any previous release, and its pricing makes it genuinely disruptive for high-volume developer workloads. The benchmarks — while self-reported at launch — are credible enough to warrant testing. At $0.30/M tokens (promotional) it is an obvious first test for any team currently paying GPT-5.5 or Claude rates.
Compare it: MiniMax M3 vs GPT-5.5 vs Claude Opus 4.8. Check the deal: MiniMax M3 50% launch discount guide.