Chinese AI startup MiniMax released the weights of its MiniMax M2.7 model on April 12. This release is part of a trend of open-weight disclosures by Chinese labs, coinciding with DeepSeek’s impending launch of its V4 model later in April.
MiniMax M2.7, initially announced in March, scored 56.22% on the SWE-Pro benchmark, matching GPT-5.3-Codex. The model achieved 55.6% on the VIBE-Pro benchmark and is described by the company as “nearly on par with Opus 4.6” for project delivery across various tasks. On the GDPval-AA benchmark, M2.7 posted an ELO of 1,495, the highest among available open-source models. The weights are now accessible on Hugging Face and supported on NVIDIA platforms.
The release faced criticism on Reddit due to its licensing terms, which restrict commercial use without prior written permission. Some community members assert that these conditions disqualify it from being truly “open source.” MiniMax characterized M2.7 as its first model to be part of its “self-evolution” development cycle.
Video: MiniMax
This development follows Zhipu AI’s release of its GLM-5.1 model on April 7. GLM-5.1, which was open-sourced under a permissive MIT license, is a 754-billion-parameter model capable of executing engineering tasks independently for up to eight hours. In contrast, Alibaba launched its proprietary API model, Qwen 3.6 Plus, on April 2, raising concerns about the company’s shift from its previous open release approach. Qwen 3.6 Plus continues to be free on OpenRouter as of April 11, though its future availability remains uncertain.
Attention is now focused on DeepSeek’s V4 model, with the company’s founder Liang Wenfeng confirming a late April release. Reuters reported that V4 will utilize Huawei’s latest Ascend chips, marking significant progress in China’s semiconductor self-sufficiency efforts. The model is anticipated to include approximately one trillion parameters using a Mixture-of-Experts architecture and a one-million-token context window. A leaked screenshot indicates V4 may launch in various modes, including a Vision mode for multimodal functionalities. Despite previous delays, recent testing of a V4-Lite version on API infrastructure signals that the public release may be imminent.
