README.md

---
base_model:
- moonshotai/Moonlight-16B-A3B
license: mit
pipeline_tag: image-text-to-text
library_name: transformers
---

<div align="center">
  <a href="Kimi-VL.pdf">KIMI-VL TECHNICAL REPORT</a>
</div>

<div align="center">
  <a href="https://arxiv.org/abs/2504.07491"><img src="figures/logo.png" height="16" width="16" style="vertical-align:middle"><b> Tech Report</b></a>  |  
  <a href="https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct"><img src="https://huggingface.co/front/assets/huggingface_logo-noborder.svg" height="16" width="16" style="vertical-align:middle"><b> HuggingFace</b>
  </a> |
  <a href="https://huggingface.co/spaces/moonshotai/Kimi-VL-A3B-Thinking/">💬 Chat Web</a>
</div>


## Introduction

We present **Kimi-VL**, an efficient open-source Mixture-of-Experts (MoE) vision-language model (VLM) that offers **advanced multimodal reasoning, long-context understanding, and strong agent capabilities**—all while activating only **2.8B** parameters in its language decoder (Kimi-VL-A3B).  Kimi-VL demonstrates strong performance across diverse challenging vision-language tasks, including college-level image and video comprehension, optical character recognition (OCR), mathematical reasoning, multi-image understanding, and more. It effectively competes with cutting-edge efficient VLMs like GPT-4o-mini, Qwen2.5-VL-7B, and Gemma-3-12B-IT, even surpassing GPT-4o in several specialized domains. Kimi-VL also excels in processing long contexts and high-resolution images, achieving impressive results on benchmarks like LongVideoBench, MMLongBench-Doc, InfoVQA, and ScreenSpot-Pro.  We also introduce **Kimi-VL-Thinking**, a variant fine-tuned for long-horizon reasoning, achieving high scores on MMMU, MathVision, and MathVista with a compact 2.8B activated LLM parameter footprint.


## Architecture

Kimi-VL uses a Mixture-of-Experts (MoE) language model, a native-resolution visual encoder (MoonViT), and an MLP projector.

<div align="center">
  <img width="90%" src="figures/arch.png">
</div>

## Model Variants

| **Model** | **#Total Params** | **#Activated Params** | **Context Length** | **Download Link** |
| :------------: | :------------: | :------------: | :------------: | :------------: |
| Kimi-VL-A3B-Instruct | 16B | 3B | 128K   | [🤗 Hugging Face](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct)   |
| Kimi-VL-A3B-Thinking  | 16B | 3B |  128K   | [🤗 Hugging Face](https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking)   |

For general multimodal tasks, OCR, long video/document understanding, video perception, and agent applications, we recommend `Kimi-VL-A3B-Instruct`. For advanced text and multimodal reasoning (e.g., math), use `Kimi-VL-A3B-Thinking`.  You can also chat with the `Kimi-VL-A3B-Thinking` model on our [Huggingface Demo](https://huggingface.co/spaces/moonshotai/Kimi-VL-A3B-Thinking/).


## Performance

Kimi-VL robustly handles diverse tasks (perception, math, college-level problems, OCR, agent interaction) across various input formats (image, multi-image, video, long-document).  See the Tech Report for detailed benchmark results.  A brief comparison with other models:

<div align="center">
  <img width="100%" src="figures/instruct_perf.png">
</div>


## Example Usage (Transformers)

```python
from PIL import Image
from transformers import AutoModelForCausalLM, AutoProcessor

model_path = "moonshotai/Kimi-VL-A3B-Instruct"
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype="auto", device_map="auto", trust_remote_code=True)
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)

image_path = "./figures/demo.png"
image = Image.open(image_path)
messages = [
    {"role": "user", "content": [{"type": "image", "image": image_path}, {"type": "text", "text": "What is the dome building in the picture? Think step by step."}]}
]
text = processor.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
inputs = processor(images=image, text=text, return_tensors="pt", padding=True, truncation=True).to(model.device)
generated_ids = model.generate(**inputs, max_new_tokens=512)
generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
response = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
print(response)
```

## Deployment (vLLM)

We have submitted a Merge Request [#16387](https://github.com/vllm-project/vllm/pull/16387) to vLLM for easier deployment.

## Citation

```
@misc{kimiteam2025kimivltechnicalreport,
      title={{Kimi-VL} Technical Report}, 
      author={Kimi Team and ...},
      year={2025},
      eprint={2504.07491},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.07491}, 
}
```