← All models

Kimi-VL-A3B-Instruct

Published Publisher Moonshot AI

Model Overview

Kimi-VL-A3B-Instruct is a vision-language instruction model designed for efficient inference. It uses a mixture-of-experts (MoE) language model, the native-resolution vision encoder MoonViT, and a projection layer. The official listing states that it has approximately 16 billion total parameters, approximately 3 billion activated parameters, and a 128K context window. (huggingface.co)

Capabilities and Uses

The model can combine text and visual information for multimodal understanding. It is suitable for image question answering, multi-image understanding, OCR, long-document and video content understanding, and agent scenarios such as operating system interfaces. The specific input processing method for video applications depends on the deployment implementation. (huggingface.co)

Limitations and Deployment

  • For more complex text or multimodal reasoning tasks, the official recommendation is to consider the Thinking version in the same series; that version's capabilities or evaluation results cannot be directly attributed to this model. (huggingface.co)
  • Model outputs may contain errors; OCR, visual recognition, and agent operation results should be verified according to application risk.
  • The official model repository provides weights, and the model card specifies an MIT license; self-deployment still incurs computing, storage, and operations and maintenance costs. (huggingface.co)

Pricing

As of 2026-11-18, checks of the official domestic and international platforms did not verify any hosted API input or output prices specifically applicable to Kimi-VL-A3B-Instruct. RMB or USD prices for other Kimi models cannot be applied to this model, nor can these findings establish whether the official provider offers a hosted service for this model. (platform.kimi.com)

Release Date and Official Sources