UI-Venus-2-9B: Ant Group's 9B GUI Agent Shipped With No Paper, No Announcement, and Near-SOTA Numbers

UI-Venus-2-9B went up on Hugging Face on August 26, 2026 with everything that usually surrounds a model launch missing. No announcement post. No press release. No technical report — the README's arXiv badge points at a placeholder URL, and the citation block carries an explicit TODO: update the arXiv identifier once the UI-Venus-2 report is public. The project-page link is a placeholder too. At the time of writing the repo has zero downloads and one like, and the four safetensors shards inside it describe a general-purpose GUI agent that, on its own model card, beats Q​wen3.6-27B on AndroidWorld and OSWorld-Verified, beats Claude Opus 4.6 on WebVoyager, beats O​penAI Operator on Online-Mind2Web, and cuts the OSBlind attack success rate from the 90%+ typical of prior GUI agents to 18.5%. InclusionAI's model cards attribute the family to the Venus Team at Ant Group, which makes this Ant Group's quiet entry into the open-weights computer-use race — and the quietest part is that none of those numbers have been independently reproduced yet.

What actually shipped

The repo is called inclusionAI/UI-Venus-2-9B, tagged image-text-to-text with a qwen3_5 architecture tag. The model is 9B parameters initialized from Q​wen3.5-9B, released under Apache-2.0 in BF16. It is a foundation GUI agent: it observes an interface, reasons about the task state, executes an action, and incorporates the resulting feedback into the next decision — a closed observe–reason–act–feedback loop that the card says is trained to span mobile applications, web platforms, and desktop operating systems in one checkpoint.

• Parameters — 9B, initialized from Q​wen3.5-9B, BF16, four shards.

• License — Apache-2.0.

• Context — the official vLLM invocation sets --max-model-len 262144 (256K tokens), which is how much screenshot history the agent can keep in one episode.

• Serving — first-class vLLM support with an O​penAI-compatible endpoint and a --reasoning-parser qwen3 flag, meaning the model emits reasoning tokens before its action, in the Q​wen3 style.

• Reach — the card claims training coverage over 170+ multilingual mobile apps (100+ Chinese, 70+ English), 50k+ live websites, and native desktop OS with 50+ professional applications.

The reasoning-parser detail matters more than it looks. A GUI agent that thinks before it acts is not new, but shipping one that plugs into standard vLLM with an O​penAI-compatible API and a stock Q​wen reasoning parser means the serving story is boring in the good way — one command and it is up, which is unusually mature for a model that was created yesterday.

The numbers, and what they are not

Everything below is vendor-reported. The model card publishes comparison tables with strong published baselines, and the row that reads best for UI-Venus-2-9B is AndroidWorld 80.2 — above Q​wen3.6-27B at 70.3, MAI-UI-8B at 70.7, and GUI-Owl-1.5-32B-Instruct at 69.8. On OSWorld-Verified the 9B posts 70.8, again ahead of Q​wen3.6-27B at 62.0 but behind Kimi K2.6 at 73.1. On web navigation the card lists WebVoyager 90.8, above Claude Opus 4.6 at 88.0 and Seed-2.0 at 85.1, and Online-Mind2Web 74.0 against O​penAI Operator's 61.3.

The most striking internal comparison is against the model's own base. On DeskCraft, UI-Venus-2-9B scores 48.0 while the Q​wen3.5-9B it was initialized from scores 14.6 — a 33.4-point jump that is the entire argument for the training recipe, and it is repeated on CAPTCHA: MCA-Bench 75.7 against the base's 30.4, NextGen-CAPTCHAs 47.6 against 2.8. The CAPTCHA numbers are worth a beat on their own, because they are the clearest sign the model was built to operate real, hostile, production websites rather than clean benchmark pages.

• Mobile — AndroidWorld 80.2, MobileWorld 65.8, VenusBench-Mobile 46.5.

• Computer use — OSWorld-Verified 70.8, DeskCraft 48.0.

• Web — WebVoyager 90.8, Online-Mind2Web 74.0, REAL 76.9, Odysseys 77.3.

• GUI grounding — ScreenSpot-Pro 73.0, VenusBench-GD 77.1, OSWorld-G-R 78.5, UI-Vision 53.2.

• CAPTCHA — Spatial-CAPTCHA-Bench 42.8, MCA-Bench 75.7, NextGen-CAPTCHAs 47.6, OpenCaptcha 50.7.

• Safety — OSBlind attack success rate 18.5%, against Claude 4.5 Sonnet at 73.0 and GPT-4o at 93.3 on the same card.

The recipe: five task families, three stages

The training pipeline the card describes is the reason the base-model deltas are so large. UI-Venus-2 is trained on a deliberately complementary mixture of five task families — Grounding, CAPTCHA, Mobile, Web, and Computer — through three stages. First, large-scale trajectory-based multimodal mid-training over simulated mobile, web, and OS environments, with human–discriminator collaborative verification filtering invalid interactions, plus programmatically synthesized grounding and CAPTCHA supervision. Second, offline step-level reinforcement learning for Mobile, OS, and Web that optimizes state-aware action selection, with verified CAPTCHA and grounding instances embedded in realistic interfaces to teach precise localization under visual clutter. Third, multi-teacher on-policy distillation that consolidates domain-specialized teachers into one unified policy.

The verification layer is the most distinctive part. Task completion is judged on task-relevant visual keypoints rather than a coarse final-screen look, and multiple models vote to aggregate heterogeneous judges — a design aimed at making the RL reward signal robust to reward hacking. Verified feedback is then distilled back into training as reflection supervision, so the agent can distinguish partial progress from true completion and recover during long-horizon interaction. None of this is independently verified, and the card gives no reproductions to point to — it is the claim the paper is supposed to back up.

The family, and the Ant Group connection

UI-Venus-2-9B is the third generation of a family inclusionAI has been shipping quietly all along. The original UI-Venus technical report appeared in August 2025 with 7B and 72B checkpoints; UI-Venus 1.5 followed in February 2026 with 2B, 8B, and 30B-A3B variants. The citation block on the new card credits the "Venus Team, Ant Group," which is how inclusionAI's relationship to Ant Group is stated in its own words — this is Ant Group's AI lab publishing under the inclusionAI name. The family's founding bet is that a GUI agent should work from pixels alone: observe the screen through screenshots rather than DOM trees, accessibility XML, or UI dumps, which is what makes it portable to the large population of Chinese and English mobile apps that expose no structured UI data. The companion Venus Framework, from the family's GitHub, drives Android automation and supports parallel multi-device control.

What we don't know yet

The honest reading of this repo is that the gap between the claims and the evidence is still wide. There is no public UI-Venus-2 technical report — the badge is a placeholder, and the README says the report is "not public" — so the benchmark tables stand as the vendor's own numbers with no methodology document and no independent rerun. There is no live project page. There is no announcement, which means no official pricing, no hosted API, and no confirmed roadmap. The card's own tables repeatedly list a UI-Venus-2-27B — 84.0 on AndroidWorld, 80.5 on OSWorld-Verified, 12.3 OSBlind — that does not exist on Hugging Face as of writing, so the family's flagship is referenced but unreleased. And with zero downloads, nobody outside the lab has run this checkpoint yet.

How you'd actually run it

For all the unknowns, the one thing that is concrete is serving. The card documents a single vLLM command that brings up the model with a 256K context window and an O​penAI-compatible API:

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

That makes the practical path self-hosting — download the weights, run vLLM, and point your agent harness at localhost. This is the case where a routing layer earns its keep on the surrounding stack rather than the GUI agent itself: the model that writes the task plan, the vision model that double-checks a screen state, the transcription model that logs what happened — all of those sit on OrcaRouter at provider list price with 0% markup and automatic failover, so you can prototype against a day-old model without betting a production pipeline on it. The day a GUI-agent checkpoint like this reaches a routed provider, it appears at the provider's list price with the same pass-through, and vendor price cuts land on our side the same day. Until then, the Apache-2.0 weights are the whole story.

What to watch next

Three things will decide whether this quiet ship turns into news. Whether the UI-Venus-2-27B that fills the card's tables actually appears on Hugging Face, and whether it carries the 12.3 OSBlind number with it. Whether the technical report lands and holds up to the first independent reproduction. And whether Ant Group decides the quiet phase is over — because a 9B open-weights agent with these table numbers and a 256K context is not something that stays at zero downloads for long.