Kimi-VL-A3B-Instruct
Kimi VL Instruct joins the Moonlight language model with a native-resolution vision encoder. Its sparse language decoder activates 2.8 billion parameters for multimodal understanding.
Images reach the language model through a projector
The system contains a MoonViT encoder, an MLP projector and a mixture-of-experts language decoder. The original report evaluates image and video comprehension, OCR, multiple images and screen-based agent tasks. Native-resolution processing is part of the vision design.Kimi VL: original technical report
Choose the generation style deliberately
Moonshot recommends Instruct for efficient general perception, OCR, long videos and documents. Its separate Thinking releases add longer reasoning. The project recommends a lower sampling temperature for Instruct than for Thinking.Kimi VL release history and model recommendations
Check the representation of a document
Practical guidance. A model that can inspect page images still needs those images prepared correctly. Check small text, page order and evidence for each extracted field. The size of the context window does not establish extraction accuracy.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | April 2025 |
| Access | Downloadable |
| Model identifiers | moonshotai/Kimi-VL-A3B-Instruct |
| Context window | 131,072 tokens Published checkpoint context limit; visual tokens and text share this budget. Kimi-VL-A3B-Instruct checkpoint configuration |
| Parameters | Approximately 16 billion total; 2.8 billion active in the language decoder.Kimi-VL-A3B-Instruct model card |
| Language model ancestry | Moonlight-16B-A3B.Kimi-VL-A3B-Instruct model card |
| Weight license | MIT.Kimi-VL-A3B-Instruct model card |
Related models
- Kimi-VL-A3B-Thinking-2506
Moonshot AI · Kimi VL
Revised visual-reasoning model with higher-resolution image input and improved video understanding.
Launch date Downloadable - Moonlight-16B-A3B
Moonshot AI · Moonlight
Pretrained language model used to study Muon optimization at 16 billion total parameters.
Launch date February 2025Downloadable - MoonViT-SO-400M
Moonshot AI · MoonViT
Standalone vision encoder adapted from SigLIP-SO-400M for native-resolution image features.
Launch date not confirmedDownloadable