Kimi-VL-A3B-Thinking-2506
Kimi VL Thinking-2506 is the June revision of the small-active-parameter visual reasoning model. It improves perception as well as reasoning, rather than only extending the length of generated thoughts.
More image detail changes the workload
The release raises the single-image pixel budget to about four times that of the original Thinking model. Moonshot reports gains in high-resolution perception, screen grounding and video reasoning. Its claim of shorter average thinking traces is a benchmark observation, not a fixed output limit.Kimi VL Thinking 2506 announcement and cookbook
Separate reasoning from the final answer
The cookbook demonstrates a 32,768-token generation budget and a parser for the model's thinking markers. It includes image, video, PDF and screenshot examples. Those examples document how to prepare inputs and read outputs from this particular release.Kimi VL Thinking 2506 announcement and cookbook
A screen coordinate is only one part of an action
Practical guidance. When this model supplies a location on a screen, check the image scale before sending a click. Keep action permissions in the application that executes the action, and verify the result after execution.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable |
| Model identifiers | moonshotai/Kimi-VL-A3B-Thinking-2506 |
| Context window | 131,072 tokens Published checkpoint context limit; visual tokens and text share this budget. Kimi-VL-A3B-Thinking-2506 checkpoint configuration |
| Training ancestry | Further-trained from Kimi-VL-A3B-Instruct.Kimi-VL-A3B-Thinking-2506 model card |
| Parameters | Approximately 16 billion total; 3 billion active in the publisher's rounded specification.Kimi VL release history and model recommendations |
| Weight license | MIT.Kimi-VL-A3B-Thinking-2506 model card |
| Image budget | Up to about 3.2 million pixels in a single image.Kimi VL Thinking 2506 announcement and cookbook |
Related models
- Kimi-K2.5
Moonshot AI · Kimi K2
Multimodal K2 successor that combines image and video understanding with coding and reasoning.
Launch date Downloadable - Kimi-VL-A3B-Instruct
Moonshot AI · Kimi VL
Vision-language model for images, documents and video, built from Moonlight and the MoonViT encoder.
Launch date April 2025Downloadable - Kimi-VL-A3B-Thinking
Moonshot AI · Kimi VL
Original long-reasoning vision-language checkpoint, retained alongside the later June 2025 revision.
Launch date April 2025Downloadable