Kimi-VL-A3B-Thinking

The first Kimi VL Thinking release adds extended visual reasoning to Kimi VL Instruct. Its weights remain published, but Moonshot recommends the separate Thinking-2506 revision for new use.

Longer reasoning without a larger decoder

The original technical report describes the Thinking model as a further-trained variant of Kimi VL. It targets long-horizon reasoning while retaining the same sparse decoder size. This is a training-stage distinction, not an increase from a small model to a larger dense model.Kimi VL: original technical report

Keep the April and June results separate

The official project marks the original Thinking release as deprecated and identifies the June checkpoint separately. That recommendation concerns which weight release to use; it does not establish that the original files have been removed.Kimi VL release history and model recommendations

Practical guidance. For a historical comparison, pin the exact checkpoint. Do not replace the original model with the June revision while keeping its old label or benchmark result.

Specifications & access

Kimi-VL-A3B-Thinking specifications
DetailValue and conditions
Launch dateApril 2025
AccessDownloadable
Model identifiersmoonshotai/Kimi-VL-A3B-Thinking
Context window131,072 tokens

Published checkpoint context limit; visual tokens and text share this budget.

Kimi-VL-A3B-Thinking checkpoint configuration
Training ancestryKimi-VL-A3B-Instruct, followed by long-chain-of-thought supervision and reinforcement learning.Kimi-VL-A3B-Thinking model card
ParametersApproximately 16 billion total; 2.8 billion active in the language decoder.Kimi-VL-A3B-Thinking model card
Weight licenseMIT.Kimi-VL-A3B-Thinking model card
Publisher recommendationSuperseded by Kimi-VL-A3B-Thinking-2506; original files remain available.Kimi-VL-A3B-Thinking model card

Related models

  • Revised visual-reasoning model with higher-resolution image input and improved video understanding.

    Launch date Downloadable
  • Vision-language model for images, documents and video, built from Moonlight and the MoonViT encoder.

    Launch date April 2025Downloadable