Purpose
Kimi-VL is an open-source Mixture-of-Experts (MoE) vision-language model (VLM) aimed at multimodal reasoning, long-context understanding, and agent capabilities. Its language decoder activates only 2.8B parameters (Kimi-VL-A3B), making it an efficient model.
Key Capabilities
- General vision-language tasks: covering fine-grained perception, mathematics, university-level questions, OCR, agents, etc., and supporting multiple input forms such as single image, multi-image, video, and long documents.
- Multi-turn agent interaction: for tasks such as OSWorld, it is reported by the project to reach a level comparable to flagship models.
- Long-context processing: equipped with a 128K extended context window, the project reports a LongVideoBench score of 64.5 and an MMLongBench-Doc score of 35.1.
- Native-resolution vision encoder MoonViT: capable of understanding ultra-high-resolution visual inputs, the project reports an InfoVQA score of 83.2 and a ScreenSpot-Pro score of 34.5.
- Long-thinking variant Kimi-VL-Thinking: developed through long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL), with the project reporting MMMU 61.7, MathVision 36.8, and MathVista 71.3.
- Kimi-VL-A3B-Thinking-2506 variant: reports higher accuracy on multimodal reasoning benchmarks (MathVision 56.9, MathVista 80.1, MMMU-Pro 46.3, MMMU 64.0), with an average thinking length reduced by 20%; it also improves general visual perception and understanding (MMBench-EN-v1.1 84.4, MMStar 70.4, RealWorldQA 70.0, MMVet 78.4); extends to video scenarios (VideoMMMU 65.2, Video-MME 71.9); supports up to 3.2 million total pixels for a single image (1792x1792), 4 times that of the original version, and reports V* Benchmark 83.2, ScreenSpot-Pro 52.8, and OSWorld-G 52.5.
Architecture
The model uses an MoE language model, a native-resolution vision encoder (MoonViT), and an MLP projector.
Model Variants and Recommended Parameters
The project provides multiple variants: Kimi-VL-A3B-Instruct (recommended for efficient inference in general multimodal perception and understanding, OCR, long video, long documents, video perception, and OS-agent), Kimi-VL-A3B-Thinking (long-thinking variant), and Kimi-VL-A3B-Thinking-2506. Recommended parameter settings: for the Thinking model, Temperature = 0.8 is suggested; for the Instruct model, Temperature = 0.2 is suggested.
Usage
- Environment: python=3.10, torch=2.5.1, transformers=4.51.3 are recommended.
- Inference can be performed based on Hugging Face Transformers. If you encounter insufficient GPU memory or want to accelerate inference, you can install flash-attn: pip install flash-attn --no-build-isolation.
- Fine-tuning: supported through LLaMA-Factory. Single-GPU LoRA fine-tuning requires 50GB of GPU memory, and multi-GPU full/lora fine-tuning (DeepSpeed ZeRO-2) is also supported.
- Deployment: the vLLM main branch already supports Kimi-VL deployment, and the vllm serve command can be used; offline inference and an OpenAI-compatible server are also supported.
- Demo: a model demo is provided on Hugging Face spaces.
Update Timeline
- 2025.06.21: Released Kimi-VL-A3B-Thinking-2506.
- 2025.04.15: vLLM supports Kimi-VL deployment (#16387).
- 2025.04.14: LLaMA-Factory supports Kimi-VL fine-tuning (#7719).
Limitations
The materials do not specify the specific license, commercial use terms, training data sources, or complete benchmark testing environment, and these are not provided for now.
