Kimi-VL-A3B-Instruct

Kimi-VL-A3B-Instruct is Moonshot AI's open vision-language model, released in April 2025 as the company's first open multimodal checkpoint. It pairs a mixture-of-experts language model (16B total, 3B activated) with the native-resolution MoonViT visual encoder and an MLP projector, handling single images, multi-image sets, long video, long documents, OCR, and OS-agent grounding across a 128K context.

The model ships under the MIT license with vLLM and LLaMA-Factory support added within days of release. A thinking variant, Kimi-VL-A3B-Thinking-2506, followed in June 2025 for stronger multimodal reasoning.