Chinese artificial intelligence company MiniMax announced its new open-weight language model, M3, on June 1, 2026. According to the company, M3 combines frontier-level coding performance with a 1-million-token context window and native multimodal capability that processes text, image and video inputs within a single model. MiniMax is positioning M3 as a cost-efficient alternative to closed-source market leaders.
In benchmark results shared by the company, M3 scored 59.0 percent on SWE-Bench Pro, 66.0 percent on Terminal-Bench 2.1 and 83.5 on BrowseComp, which measures autonomous web browsing. MiniMax says these figures place the model ahead of GPT-5.5 and Gemini 3.1 Pro on coding, and slightly past Claude Opus 4.7 on autonomous browsing. The model launched on OpenRouter at roughly $0.60 per million input tokens and $2.40 per million output tokens, with a temporary 50 percent launch discount lowering those rates to about $0.30 and $1.20.
The model is built on the company's proprietary MiniMax Sparse Attention (MSA) architecture. According to MiniMax, the design cuts per-token compute at a 1-million-token context to one-twentieth of the previous generation, while delivering more than 9 times faster prefill and more than 15 times faster decoding. The release arrives amid intense competition, as Chinese labs including DeepSeek, Qwen, Kimi and GLM continue to close the gap with frontier systems through open-weight models.
The benchmark claims, however, warrant caution. Many of the results were produced on MiniMax's own infrastructure using its own agent scaffolding, and the model weights had not been fully released at launch. Some reports note that the company stopped short of a full open-source commitment. As a result, the performance claims are unlikely to gain broad acceptance until independent verification is complete.