---
language:
- en
library_name: transformers
license: apache-2.0
pipeline_tag: image-text-to-text
tags:
- multimodal
---
# Keye-VL-2.0-30B-A3B
[](https://x.com/KwaiKeye) [](https://discord.gg/4Q6AmzxpEK) [](https://kwai-keye.github.io/) [](https://github.com/Kwai-Keye/Keye)
[](https://arxiv.org/abs/2507.01949) [](https://arxiv.org/abs/2509.01563) [](https://arxiv.org/abs/2606.10651)
[](https://huggingface.co/Kwai-Keye/Keye-VL-8B-Preview) [](https://huggingface.co/Kwai-Keye/Keye-VL-1_5-8B/) [](https://huggingface.co/Kwai-Keye/Keye-VL-2.0-30B-A3B/)
Meet Keye-VL-2.0-30B-A3B — the latest 30B-class flagship base model in the Keye series, purpose-built to push the frontier of long-video understanding and to unlock the first generation of Agent capabilities in the Keye family.
## Highlights
* **Outstanding Video Understanding and Temporal Localization**: Across five video benchmarks, Keye-VL-2.0-30B-A3B leads open-source competitors and matches or surpasses Gemini-3-Flash on temporal grounding.
* **DSA-Native Long-Context Architecture**: Sparse attention and targeted feature aggregation enable precise hour-long video understanding while keeping computation efficient.
* **High-Efficiency Inference and Training Stack**: DSA (DeepSeek Sparse Attention), ExtraIO, heterogeneous ViT-LM parallelism, activation optimization, and custom kernels reduce long-sequence prefill cost and boost training throughput.
* **Data-Centric Multimodal Pre-Training**: A carefully curated data pipeline, Keye-VL-1.5 vision encoder, and synthetic CoT data strengthen perception, OCR/chart/table understanding, and reasoning continuity.
* **Robust Post-Training for Reliable Reasoning**: MOPD, bucket advantage scaling, Context-RL, and high-SNR data filtering improve cross-modal expert merging, reduce hallucinations, and stabilize long-context decisions.
* **Agent-Ready Multimodal Capabilities**: Built-in Code, Tool, and Search agent abilities support repository tasks, API-style tool use, web-grounded search, and visual self-correction workflows.
As the first multi-modal model to land DSA in production, Keye-VL-2.0-30B-A3B delivers nearly lossless reasoning over 256K ultra-long context. It tops video understanding benchmarks at its scale and consistently rivals — or surpasses — top-tier closed-source models on fine-grained temporal perception. More importantly, it is the first Keye base model to ship with a built-in Agent collaboration mechanism, demonstrating solid system-level orchestration in Search, Tool, and Code scenarios.
## Model Performance on Benchmarks
We compare Keye-VL-2.0-30B-A3B against leading open- and closed-source models (Qwen3.5-35B-A3B, InternVL3.5-241B-A28B, GPT-5-mini, Qwen3-VL 30B-A3B / 32B / 235B-A22B) across **seven capability dimensions**: Video, Coding, Agent, Math & Reasoning, STEM, Instruction Following, and General VQA.
Selected highlights (see the technical report for the full table):
* **Fine-grained Temporal Understanding (TimeLens)**:
* Charades-TimeLens: **58.4** mIoU, on par with the strongest closed-source video baselines we tested (Gemini 3 Flash 61.19).
* ActivityNet-TimeLens: **58.5** mIoU, surpassing Gemini 3 Flash (56.95).
* QVHighlights-TimeLens: **70.1** mIoU, neck-and-neck with the top closed-source models on the official leaderboard and far ahead of Gemini 3 Flash (49.45).
* **Long-Context Scaling (VideoMME V2)**: Where most competitors degrade as the input frame count grows, our model's accuracy *increases* from **35.3%** at 64 frames to **42.4%** at 512 frames; the non-linear reasoning score climbs from 18.5 to 24.2.
* **Comprehensive Long-Video Understanding**:
* LongVideoBench: **74.1**, surpassing both Qwen3.5-35B-A3B and the much larger Qwen3-VL-235B-A22B, demonstrating strong long-video understanding at 30B scale.
At 30B scale, Keye-VL-2.0-30B-A3B not only outperforms open-source models with 200B+ parameters (e.g., Qwen3-VL-235B) on temporal understanding, but also goes head-to-head with — and in places exceeds — top closed-source giants.
## Quickstart
### Related Repositories
- SGLang (custom branch): https://github.com/Kwai-Keye/sglang/tree/keye-vl-v2-30b-release
- DeepGEMM (Keye support): https://github.com/Kwai-Keye/DeepGEMM/tree/keye_support
- EffectiveKernels: https://github.com/Kwai-Keye/EffectiveKernels
### Environment Setup
**Option 1 — Recommended: prebuilt Docker image**
```shell
docker run -it --gpus all kwaikeye/kwai-keye-vl:keye_vl_v2_30b_a3b
```
**Option 2 — Install from source**
```shell
# SGLang (custom branch)
git clone -b keye-vl-v2-30b-release https://github.com/Kwai-Keye/sglang.git
cd sglang
pip install -e python[all]
cd ..
# DeepGEMM (Keye support branch)
git clone -b keye_support https://github.com/Kwai-Keye/DeepGEMM.git
cd DeepGEMM
bash install.sh
cd ..
# EffectiveKernels
git clone https://github.com/Kwai-Keye/EffectiveKernels.git
cd EffectiveKernels
pip install -e . --no-deps --no-build-isolation
cd ..
```
### Minimal Launch (H800)
```shell
python3 -m sglang.launch_server \
--model-path=MODEL_NAME \
--tp-size=2 \
--trust-remote-code \
--mem-fraction-static=0.8
```
This is a standard SGLang service — call it with any standard OpenAI-compatible client.
### Client Usage
Below are example SGLang inference scripts for both image and video inputs.
All sampling parameters, such as `temperature`, `top_k`, and others, are provided for demonstration purposes only and should not be treated as recommended settings. Users are encouraged to experiment with and adjust these parameters based on their own needs.
For video frame-sampling related parameters, users may also customize them as needed. Specifically, `min_pixels` and `max_pixels` can be used to set the lower and upper token limits for each frame, while `video_total_pixels` can be used to limit the total token budget of the entire video input.
If `fps` is not specified, the default value is `2.0`.
#### Image Input
```python
import json
import requests
BASE_URL = "http://MASTER_NODE_IP:8000"
def generate(messages):
payload = {
"model": "",
"messages": messages,
"n": 1,
"temperature": 0.0,
"max_tokens": 256,
"top_k": 1,
"ignore_eos": False,
"skip_special_tokens": True,
}
resp = requests.post(
f"{BASE_URL}/v1/chat/completions",
headers={"Content-Type": "application/json"},
data=json.dumps(payload),
timeout=1800,
)
resp.raise_for_status()
return resp.json()
# Example: image + text
messages = [
{
"role": "user",
"content": [
{
"type": "image_url",
"image_url": {"url": "https://raw.githubusercontent.com/sgl-project/sglang/main/assets/logo.png"},
},
{"type": "text", "text": "Describe this image in detail."},
],
}
]
result = generate(messages)
print(result["choices"][0]["message"]["content"])
```
#### Video Input
```python
import json
import requests
BASE_URL = "http://MASTER_NODE_IP:8000"
def generate(messages):
payload = {
"model": "",
"messages": messages,
"n": 1,
"temperature": 0.1,
"max_tokens": 32760,
"top_p": 0.001,
"ignore_eos": False,
"skip_special_tokens": True,
}
resp = requests.post(
f"{BASE_URL}/v1/chat/completions",
headers={"Content-Type": "application/json"},
data=json.dumps(payload),
timeout=1800,
)
resp.raise_for_status()
return resp.json()
# Example: Video + text
messages = [
{
"role": "user",
"content": [
{
"type": "video_url",
"video_url": {
"url": video_url,
"preprocess_kwargs": {
"fps": 2.0,
"min_pixels": 128*28*28,
"max_pixels": 512*28*28,
"video_total_pixels":180*1024*28*28,
}
},
},
{"type": "text", "text": "Describe this video."},
],
},
]
result = generate(messages)
print(result["choices"][0]["message"]["content"])
```