Aguvis-7B
Pure-vision GUI agent from ICML 2025: a 7B model that operates phones and desktops from screenshots alone, no DOM or accessibility tree.
About
Aguvis-7B is the model behind AguVis, the ICML 2025 paper from HKU, NTU and Salesforce on unified pure-vision agents for GUI interaction. The idea is deliberately austere: the agent sees only screenshots and acts with mouse and keyboard, never touching the DOM, HTML or accessibility tree that most GUI agents lean on. That makes one model work across web, mobile and desktop, and the paper claims the first fully autonomous pure-vision GUI agent that performs without calling out to closed-source models mid-task.
Training runs in two stages, grounding (learning to point at UI elements) then planning and reasoning over trajectories, and the team open-sourced the datasets, models and training recipes. The released checkpoint is Aguvis-7B-720P on Hugging Face. Adoption is minimal so far and the model card declares no license, so this is a research artifact for now: interesting if you work on computer-use agents, not something to build a product on yet. For context, the field has since moved fast, with ByteDance's UI-TARS and Microsoft's OmniParser covering adjacent ground.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Authors report SOTA results on offline and online GUI benchmarks (OSWorld, AndroidWorld, Mind2Web-Live) vs prior methods (provider claim, ICML 2025 paper arXiv:2501.09884)
- No numeric table reproduced here; no independent replication verified
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 5/10
Scored 5/10 (metrics tier, metrics-v1). Evidence: ICML 2025 paper (PMLR v267, arXiv:2501.09884) from HKU/NTU/Salesforce authors; project page aguvis-project.github.io; HF repo xlangai/Aguvis-7B-720P created Jan 7, 2025 with 46 downloads and 9 likes (HF API, Oct 2026). Weights downloadable. Strengths: peer-reviewed at a top venue, fully open artifacts (data, models, recipes), clean pure-vision framing. Docked for: (1) near-zero adoption; (2) no license declared on the model card; (3) benchmark claims are the authors' own with no independent replication. License: none declared.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 3.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 5.0/10
- · ICML 2025 paper (PMLR v267, arXiv:2501.09884); HKU/NTU/Salesforce authors
- · HF repo xlangai/Aguvis-7B-720P created 2025-01-07; 46 downloads, 9 likes (HF API Oct 2026)
- · Datasets, models and training recipes open-sourced by the team
- · No license declared; benchmark claims author-reported; near-zero adoption