← All models

Kimi-VL-A3B-Thinking

Published Publisher Moonshot AI

Model Positioning

Kimi-VL-A3B-Thinking is a vision-language reasoning model in the Kimi-VL series, focused on complex text and multimodal reasoning. It uses a mixture-of-experts (MoE) architecture, with approximately 16 billion total parameters and 3 billion active parameters as listed by the official source, and a context window of 128K tokens. (huggingface.co)

Capabilities and Uses

  • Accepts images and text and generates text; it can be used for multi-image understanding, visual question answering, and mathematical reasoning.
  • The Kimi-VL series also supports document and video content understanding and agent-related tasks; the specific processing workflow depends on the deployment method.
  • The official source reports that this Thinking version scored 36.8, 71.3, and 61.7 on MathVision, MathVista mini, and the MMMU validation set, respectively; these are results on specific benchmarks and do not guarantee performance on real-world tasks. (huggingface.co)

Deployment and Limitations

  • The official model repository provides weights, and the model card specifies an MIT license, allowing self-deployment; self-deployment still incurs computing resource and operations costs. (huggingface.co)
  • The official source marked the original Kimi-VL-A3B-Thinking as deprecated and released a different version, Kimi-VL-A3B-Thinking-2506, on June 21, 2025; the two should not be treated as the same model. For tasks such as general OCR, long videos, and long documents, the original model card recommends the Instruct version to improve inference efficiency. Model outputs may be incorrect, and important results should be verified. (github.com)

Release Date

April 9, 2025. The commit history of the official model repository shows an initial commit and model upload records on that date; the commit date does not necessarily coincide with the time of a separate release announcement. (huggingface.co)

Pricing

As of November 18, 2026, after reviewing Moonshot's official domestic and international API pricing pages, the official API pricing for this specific model has not been verified: the charges for input and output, applicable currencies, and billing units are all unverified. Prices for other Kimi models are not applied to this model; self-deployment costs depend on the infrastructure used. (platform.kimi.com)

Official Sources