Kimi-VL-A3B-Thinking

Kimi-VL-A3B-Thinking is Moonshot AI's reasoning-tuned vision-language model, released April 10, 2025 alongside the Kimi-VL family. Built on a 16.4B-parameter Mixture-of-Experts backbone that activates only 2.8B parameters per step, it was fine-tuned with long chain-of-thought supervised learning and reinforcement learning for deep multimodal reasoning.

It competes with much larger efficient VLMs such as Qwen2.5-VL-7B and Gemma-3-12B on math and visual reasoning benchmarks like MMMU, MathVision, and MathVista (vendor-reported), and supports 128K context with the native-resolution MoonViT encoder. Everything is released under the MIT license, with code and weights on GitHub and Hugging Face; a 2506 update in June 2025 pushed accuracy further with fewer thinking tokens.