← All models

Kimi-VL-A3B-Thinking-2506

Published Publisher Moonshot AI

Model Positioning

Kimi-VL-A3B-Thinking-2506 is a multimodal vision-language reasoning model in the Kimi-VL series, intended for image understanding, long-context, and agent tasks. It uses a mixture-of-experts (MoE) language model, the native-resolution visual encoder MoonViT, and a projector. The official listing states approximately 16B total parameters, with approximately 3B activated per pass, and a context window of 128K tokens. (github.com)

Capabilities and Uses

  • Visual understanding and reasoning: Can be used for image question answering, chart understanding, mathematical reasoning, text recognition, and high-resolution image analysis. The official source states that a single image supports approximately 3.2 million pixels, such as 1792×1792. (huggingface.co)
  • Long documents and video: Can be used for long PDF understanding, video content understanding, and video reasoning; actual use requires preparing document pages or video frames according to the deployment setup. (huggingface.co)
  • Agents: Can analyze screenshots of operating system interfaces to assist in locating interface elements and generating action suggestions; actual execution requires an accompanying agent or tools. (huggingface.co)
  • Deployment: The model weights have been officially released, the model page specifies an MIT license, and usage examples are provided for Transformers, vLLM, and others. (huggingface.co)

Limitations and Considerations

The model may produce incorrect recognition results, reasoning, or action suggestions, and important results should be manually verified. Long thinking outputs and high-resolution visual inputs increase inference resource requirements; the official example sets a maximum of 32,768 output tokens for general thinking tasks, and the available length and performance depend on the deployment configuration. The official recommendation is to set the temperature to 0.8 for general thinking tasks and to 0.2 for interface element localization tasks to improve stability. (huggingface.co)

Pricing

As of 2026-11-18, the official domestic pricing page and official international pricing page had been checked, but official hosted API prices in Chinese yuan (CNY) or US dollars (USD) for this model could not be verified. Neither input nor output billing rates were verified, and they cannot be inferred from the prices of other models. Computing and other costs incurred by self-deploying the weights depend on the deployment environment. (platform.kimi.com)

Release Date and Official Sources