August 3 News.Kunlun WanweiAnnouncing the launch andOpen Sourceall-newLarge model of inference Skywork MindLink.
It is reported that this new reasoning paradigm is able to adaptively integrate reasoned and non-reasoned generated responses according to the task difficulty, which greatly reduces the reasoning cost and enhances the readability and effectiveness of the reasoning process.
The data show that the Skywork MindLink model took first place in the Humanity's Last Exam assessment out of 10 large model assessment comparisons in a non-external tool configuration, first place in four assessments in the complex math domain (Olympiad) (USAMO 2025, AIME 2024, AIME 2025, HMMT 2025), and also excelled in the general studies domain on the GPQA-diamond and MMLU-pro assessments. 2024, AIME 2025, and HMMT 2025), and excelled on the GPQA-diamond and MMLU-pro rubrics in the general studies area.
Core capabilities of the Skywork MindLink model include:
- The new reasoning paradigm Plan-based Reasoning removes the "think" label, reduces the cost of reasoning, and improves the ability to have multi-round conversations.
- A new mathematical approach is presented that can attempt to analyze the effectiveness of CoT and Plan-based Reasoning.
- Proposes a set of adaptive reasoning mechanisms capable of integrating reasoned and non-reasoned generated responses based on task difficulty.
- Post-training based on Qwen3-32B and Qwen2.5-72B improves the original model capability while saving a lot of training cost, and more size models will be open-sourced in the future.
1AI noted that currently KunlunWanwei already has big models such as Tiangong Language Big Model, Tiangong Multimodal Big Model, SWE Code Big Model, Agent Big Model, Video Big Model, 3D Big Model, Music Big Model, Audio Big Model, and so on.
Official Address:
- 72B Model weights: MindLink 72B-0801: https://huggingface.co/Skywork/MindLink-72B-0801
- Technical Report:: https://github.com/SkyworkAI/MindLink/blob/main/mindlink.pdf
- code repository:: https://github.com/SkyworkAI/MindLink