Dark Side of the Moon Open Source Multi-modal Kimi-2506

On June 23, Jin10 reported that the large model platform MoonshotAI has made significant upgrades to its Open Source multimodal model Kimi-VL-A3B-Thinking, releasing version 2506. In terms of performance, Kimi-VL-A3B-Thinking-2506 has achieved smarter and more token-efficient breakthroughs. It has attained better accuracy in multimodal reasoning benchmark tests: reaching 56.9 on MathVision (an increase of 20.1), 80.1 on MathVista (an increase of 8.4), 46.3 on MMMU-Pro (an increase of 3.2), and 64.0 on MMMU (an increase of 2.1), while the average required thinking length has decreased by 20%.