Keye-VL-1.5 is a cutting-edge Multimodal Large Language Model (MLLM) that addresses fundamental challenges in video comprehension. It features a novel Slow-Fast video encoding strategy, a progressive four-stage pre-training methodology to extend context length up to 128K tokens, and a comprehensive post-training pipeline focusing on reasoning enhancement and human preference alignment. The model demonstrates significant improvements in video understanding tasks and maintains competitive performance on general multimodal benchmarks.

[𝕏 X] [💬 Discord] [🍎 Home Page] [📖 Technical Report] [💻 GitHub Repository] [📊 Keye-VL-8B-Preview ] [📊 Keye-VL-1.5-8B ] [🚀 Demo]

🔥 News

2025.08.28 🌟 We are excited to introduce Kwai Keye-VL-1.5, a more powerful version! By incorporating innovative Slow-Fast Video Encoding strategy, new LongCoT Cold-Start data pipeline, and advanced RL training strategies, Keye-VL-1.5 reaches new heights in video understanding, image comprehension, and reasoning capabilities. Plus, it now supports an extended context length of up to 128k tokens for handling longer conversations and complex tasks. Stay tuned for more groundbreaking innovations!
2025.07.08 🌟 Keye-VL is supported by swift and vLLM. Feel free to use it without hesitation!
2025.07.03 🌟 We are excited to announce the release of our comprehensive technical report! You can read it now at arxiv.
2025.06.26 🌟 We are very proud to launch Kwai Keye-VL, a cutting-edge multimodal large language model meticulously crafted by the Kwai Keye Team at Kuaishou. As a cornerstone AI product within Kuaishou's advanced technology ecosystem, Keye excels in video understanding, visual perception, and reasoning tasks, setting new benchmarks in performance. Our team is working tirelessly to push the boundaries of what's possible, so stay tuned for more exciting updates!

Kwai Keye-VL 1.5

📐 Quick Start

Preprocess and Inference

See keye-vl-utils/README.md for details. Keye-vl-utils contains a set of helper functions for processing and integrating visual language information with Keye Series Model.

Install

pip install --upgrade keye-vl-utils==1.5.2 -i https://pypi.org/simple

Keye-VL-1.5 Inference

python

from transformers import AutoModel, AutoTokenizer, AutoProcessor
from keye_vl_utils import process_vision_info

# default: Load the model on the available device(s)
model_path = "Kwai-Keye/Keye-VL-1_5-8B"

model = AutoModel.from_pretrained(
    model_path,
    torch_dtype="auto",
    trust_remote_code=True,
    # flash_attention_2 is recommended for better performance
    attn_implementation="flash_attention_2",
).eval()

model.to("cuda")

processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)

# Image Inputs
## Non-Thinking Mode
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://s1-11508.kwimgs.com/kos/nlav11508/mllm_all/ziran_jiafeimao_11.jpg",
            },
            {"type": "text", "text": "Describe this image./no_think"},
        ],
    }
]

## Auto-Thinking Mode
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://s1-11508.kwimgs.com/kos/nlav11508/mllm_all/ziran_jiafeimao_11.jpg",
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

## Thinking mode
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://s1-11508.kwimgs.com/kos/nlav11508/mllm_all/ziran_jiafeimao_11.jpg",
            },
            {"type": "text", "text": "Describe this image./think"},
        ],
    }
]

# The default range for the number of visual tokens per image in the model is 4-20480.
# You can set min_pixels and max_pixels according to your needs, such as a token range of 32-1280, to balance performance and cost.
# min_pixels = 32 * 28 * 28
# max_pixels = 1280 * 28 * 28
# e.g.,
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://s1-11508.kwimgs.com/kos/nlav11508/mllm_all/ziran_jiafeimao_11.jpg",
                "min_pixels": 32 * 28 * 28,
                "max_pixels": 1280 * 28 * 28 
            },
            {"type": "text", "text": "Describe this image./think"},
        ],
    }
]

# Video inputs
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "video",
                "video": "http://s2-11508.kwimgs.com/kos/nlav11508/MLLM/videos_caption/98312843263.mp4",
            },
            {"type": "text", "text": "Describe this video."},
        ],
    }
]

# You can also set fps and max_frames to restrict total frames send to model.
# e.g.,

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "video",
                "video": "http://s2-11508.kwimgs.com/kos/nlav11508/MLLM/videos_caption/98312843263.mp4",
                "fps": 2.0,
                "max_frames": 1024
            },
            {"type": "text", "text": "Describe this video."},
        ],
    }
]

# Text inputs
messages = [
    {
        "role": "user",
        "content": "Hello, Keye",
    }
]

# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs, mm_processor_kwargs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
    **mm_processor_kwargs
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=1024)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

Deployment

We recommend using vLLM for fast deployment and inference.

Install

pip install keye-vl-utils==1.5.2 "vllm>=0.9.2"

Offline Inference

bash

# refer to https://github.com/QwenLM/Qwen2.5-VL?tab=readme-ov-file#inference-locally

from transformers import AutoProcessor
from vllm import LLM, SamplingParams
from keye_vl_utils import process_vision_info

model_path = "/home/keye/Keye-VL-1_5-8B"

llm = LLM(
    model=model_path,
    limit_mm_per_prompt={"image": 10, "video": 10},
    trust_remote_code=True,
)

sampling_params = SamplingParams(
    temperature=0.3,
    max_tokens=256,
)

# image
image_messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://s1-11508.kwimgs.com/kos/nlav11508/mllm_all/ziran_jiafeimao_11.jpg",
            },
            {"type": "text", "text": "Describe this image./think"},
        ],
    },
]

# video
video_messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "video",
                "video": "http://s2-11508.kwimgs.com/kos/nlav11508/MLLM/videos_caption/98312843263.mp4",
            },
            {"type": "text", "text": "Describe this video./think"},
        ],
    },
]

# Here we use video messages as a demonstration
messages = video_messages

processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
prompt = processor.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
image_inputs, video_inputs, video_kwargs = process_vision_info(
    messages, return_video_kwargs=True
)

mm_data = {}
if image_inputs is not None:
    mm_data["image"] = image_inputs
if video_inputs is not None:
    mm_data["video"] = video_inputs

llm_inputs = {
    "prompt": prompt,
    "multi_modal_data": mm_data,
    # FPS will be returned in video_kwargs
    "mm_processor_kwargs": video_kwargs,
}

outputs = llm.generate([llm_inputs], sampling_params=sampling_params)
generated_text = outputs[0].outputs[0].text

print(generated_text)

Online Serving

Serve

bash

vllm serve \
    Kwai-Keye/Keye-VL-1_5-8B \
    --tensor-parallel-size 8 \
    --enable-prefix-caching \
    --gpu-memory-utilization 0.8 \
    --host 0.0.0.0 \
    --port 8000 \
    --trust-remote-code

Openai Chat Completion Client

python

import base64
import numpy as np
from PIL import Image
from io import BytesIO
from openai import OpenAI
from keye_vl_utils import process_vision_info
import requests


# Set OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

# image url
image_messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": "https://s1-11508.kwimgs.com/kos/nlav11508/mllm_all/ziran_jiafeimao_11.jpg"
                },
            },
            {"type": "text", "text": "Describe this image./think"},
        ],
    },
]

chat_response = client.chat.completions.create(
    model="Kwai-Keye/Keye-VL-1_5-8B",
    messages=image_messages,
)
print("Chat response:", chat_response)

# image base64-encoded

import base64

image_path = "/path/to/local/image.png"
with open(image_path, "rb") as f:
    encoded_image = base64.b64encode(f.read())
encoded_image_text = encoded_image.decode("utf-8")
image_messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": f"data:image;base64,{encoded_image_text}"
                },
            },
            {"type": "text", "text": "Describe this image./think"},
        ],
    },
]

chat_response = client.chat.completions.create(
    model="Kwai-Keye/Keye-VL-1_5-8B",
    messages=image_messages,
)
print("Chat response:", chat_response)

# video, refer to https://github.com/QwenLM/Qwen2.5-VL?tab=readme-ov-file#start-an-openai-api-service
video_messages = [
    {"role": "user", "content": [
        {"type": "video", "video": "http://s2-11508.kwimgs.com/kos/nlav11508/MLLM/videos_caption/98312843263.mp4"},
        {"type": "text", "text": "Describe this video./think"}]
    },
]

def prepare_message_for_vllm(content_messages):
    vllm_messages, fps_list = [], []
    for message in content_messages:
        message_content_list = message["content"]
        if not isinstance(message_content_list, list):
            vllm_messages.append(message)
            continue

        new_content_list = []
        for part_message in message_content_list:
            if 'video' in part_message:
                video_message = [{'content': [part_message]}]\
                image_inputs, video_inputs, video_kwargs = process_vision_info(video_message, return_video_kwargs=True)
                assert video_inputs is not None, "video_inputs should not be None"
                video_input = (video_inputs.pop()).permute(0, 2, 3, 1).numpy().astype(np.uint8)
                fps_list.extend(video_kwargs.get('fps', []))

                # encode image with base64
                base64_frames = []
                for frame in video_input:
                    img = Image.fromarray(frame)
                    output_buffer = BytesIO()
                    img.save(output_buffer, format="jpeg")
                    byte_data = output_buffer.getvalue()
                    base64_str = base64.b64encode(byte_data).decode("utf-8")
                    base64_frames.append(base64_str)

                part_message = {
                    "type": "video_url",
                    "video_url": {"url": f"data:video/jpeg;base64,{','.join(base64_frames)}"}
                }
            new_content_list.append(part_message)
        message["content"] = new_content_list
        vllm_messages.append(message)
    return vllm_messages, {'fps': fps_list}


video_messages, video_kwargs = prepare_message_for_vllm(video_messages)


chat_response = client.chat.completions.create(
    model="Kwai-Keye/Keye-VL-1_5-8B",
    messages=video_messages,
    max_tokens=128,
    extra_body={
        "mm_processor_kwargs": video_kwargs
    }
)

print("Chat response:", chat_response)

Evaluation

See evaluation/KC-MMBench/README.md for details.

👀 Architecture and Training Strategy

The Kwai Keye-VL-1.5 model architecture is based on the Qwen3-8B language model and incorporates a vision encoder initialized from the open-source SigLIP. It supports SlowFast video encoding and native dynamic resolution, preserving the original aspect ratio of images by dividing each into a 14x14 patch sequence. A simple MLP layer then maps and merges the visual tokens. The model uses 3D RoPE for unified processing of text, image, and video information.

🌟 Pre-Train

A SlowFast video encoding demonstration: the Slow processes a smaller number of frames at higher resolution, while the Fast handles more frames at lower resolution.

🌟 Post-Train

The post-training process includes non-reasoning stage and reasoning stage. The non-reasoning stage is composed of SFT and MPO training. The reasoning stage consists of three key steps: CoT Cold Start (we construct a five-step construction pipeline to generate high-quality CoT Cold-Start Dataset and apply model merging to refine model performance), General RL (we concentrate on improving Keye-VL-1.5's reasoning ability, applying GSPO, we propose progressive hint sampling to fully take advantage of hard problems and iteratively improve the cold-start and general RL model), and Alignment RL (improving Keye-VL-1.5's instruction following, format adherence, preference alignment and RAG ability with our reward system, we construct instruction following data, reasoning data and RAG data for RL training in this step).

📈 Experimental Results

Keye-VL-1.5-8B establishes itself with powerful, state-of-the-art perceptual abilities that are competitive with leading models.
Keye-VL-1.5-8B demonstrates exceptional proficiency in video understanding. Across a comprehensive suite of authoritative public video benchmarks, including Video-MME, Video-MMMU, TempCompass, LongVideoBench, and MMVU, the model's performance significantly surpasses that of other top-tier models of a comparable size.
In evaluation sets that require complex logical reasoning and mathematical problem-solving, such as WeMath, MathVerse, and LogicVista, Kwai Keye-VL-1.5-8B displays a strong performance curve. This highlights its advanced capacity for logical deduction and solving complex quantitative problems.

✒️ Citation

If you find our work helpful for your research, please consider citing our work.

bibtex

@misc{Keye-VL-1.5,
      title={Kwai Keye-VL-1.5 Technical Report}, 
      author={Kwai Keye Team},
      year={2025},
      eprint={TBD},
}
@misc{kwaikeyeteam2025kwaikeyevltechnicalreport,
      title={Kwai Keye-VL Technical Report}, 
      author={Kwai Keye Team},
      year={2025},
      eprint={2507.01949},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.01949}, 
}

Acknowledgement

Kwai Keye-VL is developed based on the codebases of the following projects: SigLIP, Qwen3, Qwen2.5-VL, VLMEvalKit. We sincerely thank these projects for their outstanding work.

Kwai-Keye/Keye-VL-1_5-8B

Kwai Keye-VL 1.5

🔥 News

Contents

📐 Quick Start

Preprocess and Inference

Install

Keye-VL-1.5 Inference

Deployment

Install

Offline Inference

Online Serving

Evaluation

👀 Architecture and Training Strategy

🌟 Pre-Train

🌟 Post-Train

📈 Experimental Results

✒️ Citation

Acknowledgement

Kwai-Keye/
Keye-VL-1_5-8B