Kimi-VL-A3B-Thinking

// Overview

Kimi-VL-A3B-Thinking is the long-reasoning variant of Moonshot AI's Kimi-VL family, an open-weight Mixture-of-Experts vision-language model released in April 2025. Despite a roughly 16-billion-parameter footprint, it activates only about 2.8B parameters in its language decoder per step, pairing that efficient MoE backbone with the native-resolution MoonViT visual encoder and an MLP projector. The result is a small model that reads images, multi-image sets, video frames and long documents while keeping inference cheap.

Where the base Kimi-VL-A3B-Instruct focuses on perception, Kimi-VL-A3B-Thinking is tuned for step-by-step multimodal reasoning. The Kimi Team built it through long chain-of-thought supervised fine-tuning followed by reinforcement learning, so it spends extra tokens working through math diagrams, charts and visual logic before answering. On its release the model scored 61.7 on MMMU (val), 36.8 on MathVision and 71.3 on MathVista-mini at the compact 2.8B activated size, competitive with much larger vision models of its era.

Kimi-VL-A3B-Thinking ships under the permissive MIT license with full open weights on Hugging Face, a 128K-token context window, and a Hugging Face Spaces demo. Because it is genuinely open, it runs locally or on hosted endpoints such as Replicate, making it a practical pick for builders who need an inexpensive, auditable visual reasoner rather than a closed API. Note that Moonshot later shipped an upgraded Kimi-VL-A3B-Thinking-2506 checkpoint; this page covers the original April 2025 release.

// Benchmarks

1. MMMU (val, Pass@1)61.7%
2. MathVista-mini (Pass@1)71.3%
3. MathVision (full, Pass@1)36.8%
4. InfoVQA83.2%
5. LongVideoBench64.5%
6. MMLongBench-Doc35.1%
7. ScreenSpot-Pro34.5%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Strong multimodal chain-of-thought reasoning at a tiny ~2.8B activated-parameter cost
• Fully open weights under the permissive MIT license — local or self-hosted deployment
• Native-resolution MoonViT encoder handles high-resolution images, charts and documents
• 128K-token context for long documents, multi-image inputs and video frames
• Efficient MoE design keeps inference cost and latency low versus dense vision models

// Best for

• Visual math and diagram reasoning (geometry, charts, plots)
• Document, OCR and long-PDF understanding over long context
• Multi-image and video-frame question answering
• On-device or self-hosted multimodal apps where closed APIs aren't an option
• Research and fine-tuning baselines for efficient MoE vision-language models

// How to access

// Kimi-VL — every version

The full lineage of the Kimi-VL line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.