# Kimi-VL-A3B-Instruct

> Moonshot AI's lightweight open-weights MoE vision-language model — 16B total, 2.8B active, 128K context

- Maker: Moonshot AI (Kimi)
- Line: Kimi-VL
- Canonical page: https://ai-tldr.dev/models/kimi-vl-a3b-instruct/
- Updated: 2026-06-18

## In short

Kimi-VL-A3B-Instruct is the instruction-tuned vision-language model in Moonshot AI's (Kimi) Kimi-VL line, released in April 2025 under an MIT license with fully open weights. It has a 128K-token context window.

## Facts

| Fact | Value | Source |
| --- | --- | --- |
| Context window | 128K | [1] |
| Licence | MIT · open weights | [1] |
| Released | April 2025 | [2] |
| Parameters | 16B total / 2.8B activated (MoE, 8 of 384 experts) | [1] |

## Sources

1. [Hugging Face model card](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct)
2. [GitHub repository](https://github.com/MoonshotAI/Kimi-VL)

## Overview

Kimi-VL-A3B-Instruct is the instruction-tuned vision-language model in Moonshot AI's (Kimi) Kimi-VL line, released in April 2025 under an MIT license with fully open weights. It pairs a native-resolution visual encoder called MoonViT with a sparse Mixture-of-Experts language decoder: although the model totals 16B parameters, it activates only about 2.8B per token by routing through 8 of 384 experts, giving it the efficiency of a roughly 3B-class dense model while retaining the capacity of a much larger one.

The model handles text, single and multiple images, video sequences, and long documents within a 128K-token context window, making it well suited to OCR, college-level image and video comprehension, mathematical reasoning, and multi-image understanding. Moonshot AI positions Kimi-VL-A3B-Instruct as a compact, open alternative that competes with efficient closed and open VLMs such as GPT-4o-mini, Qwen2.5-VL-7B and Gemma-3-12B-IT, while surpassing GPT-4o in several specialized domains like document and OCR tasks.

Kimi-VL-A3B-Instruct is distributed primarily as open weights on Hugging Face for self-hosting rather than as a paid hosted API. A separate reasoning-focused sibling, Kimi-VL-A3B-Thinking, adds chain-of-thought and reinforcement-learning training on top of the same A3B backbone.

## Benchmarks

- MMMU (Val): 57% ([source](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct))
- MathVista: 68.7% ([source](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct))
- MMBench-EN-v1.1: 83.1% ([source](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct))
- AI2D: 84.9% ([source](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct))
- InfoVQA: 83.2% ([source](https://arxiv.org/abs/2504.07491))
- OCRBench: 86.7% (867/1000) ([source](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct))
- ScreenSpot-V2: 92.8% ([source](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct))
- ScreenSpot-Pro: 34.5% ([source](https://arxiv.org/abs/2504.07491))
- LongVideoBench: 64.5% ([source](https://arxiv.org/abs/2504.07491))
- MMLongBench-Doc: 35.1% ([source](https://arxiv.org/abs/2504.07491))
- VideoMMMU: 52.6% ([source](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct))

## Strengths

- Strong document understanding and OCR for its size — 86.7% on OCRBench and 83.2 on InfoVQA
- Efficient MoE design: ~2.8B active parameters keep inference cheap while 16B total capacity preserves quality
- 128K context window enables long documents, multi-image sets and video clips in a single prompt
- Capable GUI/agent grounding — 92.8 on ScreenSpot-V2 for on-screen element localization
- Fully open weights under a permissive MIT license, so it can be self-hosted and fine-tuned freely
- Solid math and college-level multimodal reasoning (68.7 MathVista, 57.0 MMMU) despite its small active footprint

## Best for

- OCR and document parsing across PDFs, scans and screenshots
- Visual question answering over charts, diagrams and infographics
- Long-document and multi-image analysis using the 128K context
- Video comprehension and summarization
- GUI agents and on-screen element grounding for automation
- Self-hosted multimodal applications where open weights and low inference cost matter

## Kimi-VL versions

- [Kimi-VL-A3B-Thinking-2506](https://ai-tldr.dev/models/kimi-vl-a3b-thinking-2506/) (21 June 2025)
- [Kimi-VL-A3B-Thinking](https://ai-tldr.dev/models/kimi-vl-a3b-thinking/) (April 2025)
- Kimi-VL-A3B-Instruct (April 2025) — this page

## FAQ

### What is Kimi-VL-A3B-Instruct?

It is Moonshot AI's (Kimi) instruction-tuned vision-language model, released in April 2025. It uses a Mixture-of-Experts design with 16B total parameters but only about 2.8B activated per token, pairing a MoonViT visual encoder with a sparse MoE language decoder for text, image, video and document understanding.

### Is Kimi-VL-A3B-Instruct open source?

Yes. The weights are openly published on Hugging Face under the permissive MIT license, so it can be self-hosted, fine-tuned and used commercially. There is no official paid hosted API for the Instruct variant — it is designed to run on your own hardware.

### What context window does Kimi-VL-A3B-Instruct support?

It supports a 128K-token context window, large enough to handle long documents, multiple high-resolution images, or lengthy video clips alongside extensive text in a single prompt.

### How does Kimi-VL-A3B-Instruct compare to other models?

Despite activating only about 2.8B parameters, it competes with efficient VLMs like GPT-4o-mini, Qwen2.5-VL-7B and Gemma-3-12B-IT, and surpasses GPT-4o in some specialized areas such as OCR (86.7 on OCRBench) and document understanding.

## Links

- [Hugging Face model card](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct)
- [GitHub repository](https://github.com/MoonshotAI/Kimi-VL)
- [Kimi-VL Technical Report (arXiv)](https://arxiv.org/abs/2504.07491)

Source: AI/TLDR LLM registry, https://ai-tldr.dev/models/kimi-vl-a3b-instruct/