Instructions to use inclusionAI/UI-Venus-2-9B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use inclusionAI/UI-Venus-2-9B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="inclusionAI/UI-Venus-2-9B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("inclusionAI/UI-Venus-2-9B")
model = AutoModelForMultimodalLM.from_pretrained("inclusionAI/UI-Venus-2-9B", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use inclusionAI/UI-Venus-2-9B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "inclusionAI/UI-Venus-2-9B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "inclusionAI/UI-Venus-2-9B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/inclusionAI/UI-Venus-2-9B

SGLang

How to use inclusionAI/UI-Venus-2-9B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "inclusionAI/UI-Venus-2-9B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "inclusionAI/UI-Venus-2-9B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "inclusionAI/UI-Venus-2-9B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "inclusionAI/UI-Venus-2-9B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use inclusionAI/UI-Venus-2-9B with Docker Model Runner:
```
docker model run hf.co/inclusionAI/UI-Venus-2-9B
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

UI-Venus-2-9B

UI-Venus-2 is a general-purpose foundation GUI agent designed to operate across mobile applications, web platforms, and desktop operating systems through a unified closed-loop reasoning–action framework: the agent observes the current interface, reasons about the task state, executes an action, and incorporates environmental feedback into its next decision.

To bridge the gap from benchmark-oriented models toward dependable real-world applications, UI-Venus-2 jointly scales three critical dimensions:

Environments — expanding the executable mobile pool to 170+ multilingual apps (100+ Chinese and 70+ English), alongside a curated web pool of 4,000+ domains across 19 categories and native desktop operating systems;
Tasks — a deep-research pipeline that grounds generated instructions in the actual functionality of each application, improving the accuracy, validity, and executability of task queries;
Verification — trace-level and sample-level evaluators based on task-relevant visual keypoints and multi-model voting, providing reliable RL reward signals that are robust to reward hacking.

UI-Venus-2 is evaluated for consequential-action safety on OSHarm and OSBlind. UI-Venus-2-9B achieves an Attack Success Rate (ASR, lower is better) of 11.3% on OSHarm and 48.8% on OSBlind, compared with 25.3% and 79.4%, respectively, for its Qwen3.5-9B base model.

This repository releases the full-parameter weights of UI-Venus-2-9B, initialized from Qwen3.5-9B, which achieves near state-of-the-art performance among models of comparable scale across GUI grounding, mobile, web, computer-use, and CAPTCHA benchmarks.

Performance overview of UI-Venus-2 across GUI agent benchmarks

Highlights

Scaled multilingual mobile-use environments with reliable task generation. A substantially expanded executable mobile pool covering Chinese and English app ecosystems, paired with a deep-research-driven query-generation strategy grounded in real application functionality.
Computer-use capability built from the ground up. Dedicated desktop operating-system capabilities constructed from scratch through computer-use data collection and task-specific training, extending the UI-Venus family to mobile, web, and OS interaction in one unified end-to-end agent.
Keypoint-grounded verification. Task completion is judged on task-relevant visual keypoints rather than a coarse holistic look at the final screen, with multi-model voting aggregating heterogeneous judges — reducing single-judge bias and making the reward signal robust to reward hacking.
Multi-teacher on-policy distillation. Domain-specialized teachers are consolidated with structured, action-aware supervision that concentrates the distillation signal on the executable behaviors governing environment interaction.

Training Pipeline

UI-Venus-2 is trained on a deliberately complementary mixture of five task families — Grounding, CAPTCHA, Mobile, Web, and Computer — through three stages:

Multimodal Mid-Training — large-scale trajectory-based training over simulated mobile, web, and OS environments, with navigation data as the dominant component and human–discriminator collaborative verification filtering invalid or ambiguous interactions.
Offline RL — independent step-level optimization across Grounding, CAPTCHA, Mobile, Web, and Computer tasks develops domain-specialized policies.
Multi-teacher On-policy Distillation — domain-specialized teachers are consolidated into a single unified policy that preserves the broad multimodal reasoning of the base model while composing spatial grounding, verified interaction, and long-horizon navigation.

The three-stage training pipeline of UI-Venus-2

The training data is produced by a deep-research-driven task-generation pipeline spanning mobile, browser, and computer environments. The active report describes a curated web pool of 4,000+ domains across 19 categories, seeded with 45,000 tasks selected from InSTA-150k-v3; synthetic grounding and CAPTCHA records are exported for task-specific SFT and RL.

System overview: task generation pipeline and trajectory collection across mobile, browser, and computer environments

Quick Start

Serving with vLLM

The following is the intended serving configuration. It was not live-canary-validated as part of this model-card update, so pin and verify a vLLM version appropriate for Qwen3.5 before deployment.

pip install -U vllm

vllm serve inclusionAI/UI-Venus-2-9B \
    --port 8000 \
    --tensor-parallel-size 1 \
    --max-model-len 262144 \
    --reasoning-parser qwen3

The server exposes an OpenAI-compatible API at http://localhost:8000/v1.

Inference Configuration

The evaluation settings depend on the task family:

General agentic tasks: use temperature 1.0, enable reasoning, and retain the complete reasoning history in the multi-step context.
GUI grounding: disable reasoning and use temperature 0 for direct point localization.
CAPTCHA: use the CAPTCHA-specific prompt and a parser/executor that supports multiple actions in one model turn.

Starting the vLLM server alone does not provide the complete closed-loop GUI agent. Reference prompts, action parsers, and runnable examples for Mobile, Computer, Browser, Grounding, and CAPTCHA are available in the UI-Venus code repository. Environment coverage differs by domain: the included Mobile and Computer examples support multi-turn inference over prerecorded screenshots, while real-device ADB automation, browser interaction, and benchmark-specific setup are documented separately in the repository.

Release Status

Model weights: included in this repository.
Code: available in the UI-Venus GitHub repository.
Project page: available at ui-venus.github.io/UI-Venus-2.
Technical report and UI-Venus-2 citation: will be added when the report is publicly released.
License: the model-weight license is pending final confirmation and will be added before public release. The Apache-2.0 declaration has intentionally not been carried over because the current upstream materials contain conflicting license statements.

Results

The tables below compare the released 9B checkpoint and the 27B family checkpoint with a representative subset of general-purpose and GUI-specialized baselines from the technical report. – denotes an unreported result, and * denotes a baseline evaluated or reproduced by the UI-Venus authors under the stated protocol. Full comparisons, source citations, and metric definitions are provided in the technical report. Live-environment results may vary with evaluation date and environment state.

Mobile Use

Reported score (%) across interactive mobile-agent benchmarks, including controlled simulations and real-application environments.

Model	MobileGym	VenusBench-Mobile	AndroidWorld	MobileWorld	KnowUBench	MemGUI
Qwen3.5-9B	9.0*	15.3*	57.8	18.0 (18.0)*	33.3	6.2*
Qwen3.6-27B	24.6*	28.0*	70.3	36.8 (41.9)*	–	25.7*
Seed-2.0-Pro	52.0	20.1*	–	63.2	51.6	65.6*
UI-Venus-1.5-30B-A3B	21.5*	21.5	77.6	17.1	–	10.9*
Qwen-UI-Agent-27B	–	–	–	82.1 (85.5)	–	–
UI-Venus-2-9B (released)	52.7	46.5	80.2	65.8 (75.2)	56.5	62.6
UI-Venus-2-27B	60.5	48.7	84.0	76.1 (82.9)	59.7	70.3

MobileWorld reports GUI-only success rate on 117 tasks: the primary value uses a 50-step budget and the value in parentheses uses 100 steps. VenusBench-Mobile uses its 149-task primary pool. MemGUI reports Main Results pass@1.

Computer Use

Reported score (%) on open-ended tasks in desktop environments; benchmark-specific definitions follow the table.

Model	OSWorld-Verified	OSWorld 2.0 Binary Acc.	OSWorld 2.0 Partial Score	DeskCraft
Qwen3.5-9B	41.8	0.0*	2.5*	14.6*
Qwen3.6-27B	62.0	0.0*	3.8*	28.7*
Kimi-K2.6	73.1	1.9	7.1	41.4*
GPT-5.5	78.7	13.0	46.7	–
Claude-Opus-4.8	83.4	–	–	–
Qwen-UI-Agent-27B	79.5	–	–	–
UI-Venus-2-9B (released)	70.8	0.0	7.5	48.0
UI-Venus-2-27B	80.5	2.8	13.2	55.5

OSWorld-Verified comparisons use the 361-task setting reported by their cited sources and may use model-specific action scaffolds, so they should be interpreted as benchmark-level references rather than controlled ablations. OSWorld 2.0 reports the official Binary Accuracy and Partial Score metrics over 108 tasks under a 150-step budget. DeskCraft is an author-evaluated aggregate over the 538-task union of the Standard and Interactive splits and differs from official split-level reporting.

Web Navigation

Success rate (%) on end-to-end web navigation; Odysseys reports the average rubric score (%).

Model	WebVoyager	Online-Mind2Web	REAL	Odysseys Avg.	Odysseys Perfect
Qwen3.5-9B	46.9*	27.3*	18.2*	42.6*	13.5*
Qwen3.6-27B	84.3*	55.3*	27.3*	39.5*	18.5*
Seed-2.0-Pro	85.1*	68.5*	74.4*	60.2*	30.1*
Claude-Opus-4.6	88.0	–	–	68.9	44.5
Fara1.5-27B	89.3	72.3	–	–	–
UI-Venus-2-9B (released)	90.8	74.0	76.9	77.3	62.0
UI-Venus-2-27B	93.4	78.3	80.2	80.4	66.3

WebVoyager uses the refreshed 595-task split; live-site states may vary by evaluation date. Odysseys evaluates 200 live-web tasks with gemini-3.1-flash-lite-preview as the rubric judge and reports both the average fraction of satisfied rubric items and the percentage of tasks for which every rubric item is satisfied.

GUI Grounding

Accuracy (%) on element localization across high-resolution professional software, mobile / web / desktop interfaces, and reasoning-heavy instructions.

Model	ScreenSpot-Pro	VenusBench-GD	OSWorld-G-R	UI-Vision
Qwen3.6-27B	65.2*	67.7*	76.9*	58.3*
Qwen 3.7 Plus	68.9	75.2*	78.2	68.0
Seed 2.1 Pro	65.3	73.9*	78.0	62.0
UI-Venus-1.5-30B-A3B	69.6	75.0	76.4	54.7
Qwen-UI-Agent-27B	76.6	–	78.5	70.0
UI-Venus-2-9B (released)	73.0	77.1	78.5	53.2
UI-Venus-2-27B	74.1	80.1	79.1	66.9

VenusBench-GD reports English-instruction micro-average point-in-box accuracy.

CAPTCHA Solving

Pass@1 (%) across five CAPTCHA benchmarks. VenusBench-CAPTCHA spans eight real-world interaction types—including slider puzzles, image rotation, text and icon click, OCR entry, and visual reasoning—and requires the full multi-action sequence in a single turn.

Model	VenusBench-CAPTCHA	MCA-Bench	Spatial-CAPTCHA-Bench	NextGen-CAPTCHAs	Open CaptchaWorld
Qwen3.5-9B	28.3	30.4	4.9	2.8	36.4
Qwen3.6-27B	53.0	51.7	31.0	14.1	47.7
Seed-2.0-Pro	47.9	36.5	43.8	20.4	55.6
Kimi-K2.6	39.7	38.7	24.8	7.2	47.8
UI-Venus-2-9B (released)	78.1	75.7	42.8	47.6	50.7
UI-Venus-2-27B	79.9	79.6	48.6	54.5	56.3

All values are Pass@1. VenusBench-CAPTCHA reports the micro-average over all 219 fixed examples. The public-benchmark evaluation uses 1,000 sampled MCA-Bench examples, 15 retained NextGen-CAPTCHAs task types, and 16 retained Open CaptchaWorld task types.

Safety

OSHarm covers explicit safety threats including deliberate misuse, prompt injection, and model misbehavior. OSBlind tests benign-looking instructions whose execution context can produce harmful outcomes. Both report Attack Success Rate (ASR, lower is better).

Model	OSHarm ASR (%) ↓	OSBlind ASR (%) ↓
Qwen3.5-9B	25.3	79.4
Qwen3.5-27B	18.0	89.3
Kimi-K2.6	32.0	93.6
EvoCUA-8B	39.3	85.3
UI-TARS-1.5	36.0	83.3
ScaleCUA	25.3	84.7
UI-Venus-2-9B (released)	11.3	48.8
UI-Venus-2-27B	15.3	47.9

These results do not make harmful-action risk negligible. Deployments should keep human confirmation and application-level safeguards around consequential operations, and CAPTCHA capabilities should only be used in authorized environments.

Citation

The UI-Venus-2 citation will be added when the technical report is public. For prior UI-Venus releases, please use:

@article{uivenus15,
  title   = {UI-Venus-1.5 Technical Report},
  author  = {{Venus Team, Ant Group}},
  journal = {arXiv preprint arXiv:2602.09082},
  year    = {2026}
}

@article{uivenus2025,
  title   = {UI-Venus Technical Report: Building High-performance UI Agents with RFT},
  author  = {{Venus Team, Ant Group}},
  journal = {arXiv preprint arXiv:2508.10833},
  year    = {2025}
}