Kunlun Tech officially released and open-sourced its latest reasoning large model Skywork MindLink, while announcing the 72B model weights, technical report, and code repository links, drawing widespread attention in the industry.
The model achieves dynamic path selection through an innovative reasoning framework, flexibly combining reasoning and non-reasoning generation modes based on task complexity, reducing computational costs while improving answer transparency and efficiency.
In benchmark tests, Skywork MindLink performed outstandingly. It not only won the championship in the "Human Final Exam" project among 10 large model evaluations without relying on external tools, but also won four gold medals in the USAMO 2025, AIME 2024/2025, and HMMT 2025 competitions in the math olympiad field. Additionally, it demonstrated excellent performance in general knowledge evaluations such as GPQA-diamond and MMLU-pro.
On the technical level, the model adopts a new paradigm called "Plan-based Reasoning," discarding the traditional "think" tag design to optimize multi-turn dialogue experiences; it introduces mathematical tools for quantitative analysis of the effectiveness of "chain-of-thought" and "plan-based reasoning" modes; it includes an adaptive reasoning system that automatically adjusts generation strategies based on task difficulty; by using post-training strategies based on the Qwen3-32B and Qwen2.5-72B models, it significantly enhances the basic model capabilities while controlling training costs.
As the latest member of Kunlun Tech's large model matrix, the release of Skywork MindLink further solidifies the company's layout in specialized large models such as language, multimodal, code, Agent, and audio-video. The company stated that it will continue to promote the open-source plan for models of different scales.
Project Address: https://github.com/SkyworkAI/MindLink