Kimi-VL-A3B-Thinking (Moonshot AI): a small open vision model with chain-of-thought reasoning
In one sentence Moonshot AI releases Kimi-VL, an open MoE vision-language model (16B total, about 3B active, MIT license), with a Thinking variant trained with long chain-of-thought and reinforcement learning. It beats GPT-4o on visual math benchmarks but trails it on MMMU.
Moonshot AI, a Chinese company, publishes Kimi-VL: a model that understands images and text. It is a "mixture of experts" model: 16 billion parameters in total but only about 3 used per token, so it is cheap to run.
The Thinking variant writes a long chain of reasoning before answering. It was trained first on step-by-step reasoning examples and then with reinforcement learning.
The result: on math problems with figures (MathVista, MathVision) it does better than GPT-4o, while on multidisciplinary university-level questions (MMMU) it stays below. Weights are downloadable under the MIT license.
Companies
Moonshot AI
Tools
Kimi-VL-A3B-Thinking, Kimi-VL-A3B-Instruct
Tags
Sources