A generated title card reading 'LingSante-3.1-air-VL vs UI-Venus-2-9B' with the subtitle 'One shared benchmark, two different jobs'. Three icon cards read 'AndroidWorld 86.84 against 80.2', 'Same lab, different training' and '250 GB against four shards'. A footer reads 'Vendor-reported scores from both cards, 23 November 2026.' The OrcaRouter logo is composited in a white strip along the bottom-right edge.
Engineering & Research

LingSante-3.1-air-VL vs UI-Venus-2-9B: One Shared Row

Autore

jinhao.song

Data di pubblicazione

Ultimi modelli · 20Vedi tutti i modelli →
Benchmark: Artificial Analysis · aggiornato ogni giorno
Torna a tutti gli articoli

Two checkpoints from the same organisation, uploaded three months apart, and exactly one benchmark in common. LingSante-3.1-air-VL is inclusionAI's 124-billion-parameter medical vision model, uploaded on 19 November 2026. UI-Venus-2-9B is the same team's 9-billion-parameter GUI agent, uploaded on 26 August 2026 and initialised from Qwen3.5-9B. Both report AndroidWorld. The medical model scores 86.84; the GUI specialist scores 80.2. That is the sort of result that produces a headline and, on closer reading, a lesson about what benchmark numbers can carry.

The headline is real but narrow. LingSante is thirteen times the stored parameters, activates 5.5B against UI-Venus's 9B dense, and needs four 141 GB accelerators against a single-GPU deployment. It was not built to drive Android apps; it was built to read chest radiographs and write reports, and GUI control is one of four secondary clinical workflows in its card — booking an appointment, buying a medicine, handling a pop-up. The fact that it edges the dedicated agent on the dedicated agent's own headline benchmark is either impressive or an artefact of how that benchmark is run. This piece argues it is mostly the second, and that the first is still worth knowing.

Same lab, different jobs

The shared provenance is not a coincidence of naming. Both repositories sit under inclusionAI, and the model cards attribute the UI-Venus family to the Venus Team at Ant Group. What differs is the training target, and the cards are unusually explicit about it.

• Parameters — LingSante-3.1-air-VL: 124B total, 5.5B active, 512 routed experts over 42 layers. UI-Venus-2-9B: 9B dense, every parameter active, initialised from Qwen3.5-9B.

• Weights on disk — 249.7 GB of bfloat16 across fifty shards versus four shards for the 9B model.

• Context — 262,144 tokens on both, though they use it differently: screenshot history for the agent, multi-image clinical records for the medical model.

• Serving — LingSante: an SGLang day-zero image with tensor parallelism 4, or an inclusionAI vLLM fork. UI-Venus: stock vLLM with an OpenAI-compatible endpoint and a Qwen3 reasoning parser.

• Licence — MIT for the medical model, with a research-only notice in prose. The UI-Venus card declares no licence field at all, which leaves those weights technically all rights reserved.

• Training reach — UI-Venus claims 170-plus multilingual mobile apps, a curated web pool of 4,000-plus domains across 19 categories, and native desktop operating systems. LingSante claims 120-plus skin conditions, four clinical focus areas, and GUI workflows as a supporting capability.

Read as a pair, they look like a deliberate division of labour: a large multimodal reasoner for clinical content, a small computer-use policy for interfaces. The odd part is that the large one also reports the small one's benchmark, and does well on it.

The one row they share

Screenshot of the left content column of the Hugging Face model card for inclusionAI/UI-Venus-2-9B captured 23 November 2026 in English: the introduction describes UI-Venus-2 as a general-purpose GUI agent operating across mobile apps, web platforms and desktop operating systems through an observe-reason-act loop; bullets describe 170-plus multilingual mobile apps, a curated web pool of 4,000-plus domains across 19 categories, and visual-keypoint verification with multi-model voting; a safety paragraph reports an attack success rate of 11.3% on OSHarm and 48.8% on OSBlind against 25.3% and 79.4% for the Qwen3.5-9B base; a final paragraph states the repository releases full-parameter weights initialised from Qwen3.5-9B.

AndroidWorld is the common ground. UI-Venus-2-9B posts 80.2, above Qwen3.6-27B at 70.3, MAI-UI-8B at 70.7 and GUI-Owl-1.5-32B-Instruct at 69.8 on its own card. LingSante-3.1-air-VL posts 86.84 in a column that also carries Kimi-K3.1-Pro at 86.20, GPT-6.1 Sol at 92.67 and Gemini 3.8 Flash at 75.86. Both figures are vendor-reported. Neither has been reproduced by a third party.

On its face the comparison says the medical model is a better Android agent than the Android agent. That conclusion deserves suspicion for a structural reason: an AndroidWorld score is not a property of weights alone. It is a property of the weights plus the agent scaffold — how the model is prompted, how screenshots are encoded, how actions are parsed, how many steps it gets, how failures are retried. Two labs reporting the same benchmark can be measuring two different systems. inclusionAI is one lab, which removes some of that variance and none of it in principle, because the two cards do not describe the same harness in the same detail. LingSante's card notes that search, tool use and GUI workflows require an external environment that maintains state and validates actions; UI-Venus's card describes an observe–reason–act–feedback loop trained across mobile, web and desktop.

So the fair reading is narrower than the number. On inclusionAI's own evaluations, a 124B medical model that treats GUI control as a secondary skill scored higher on AndroidWorld than the 9B model built for it. Whether that transfers to your app depends on a scaffold neither card publishes.

Why an AndroidWorld score travels badly

Agent benchmarks are the least portable numbers in the multimodal field, and it is worth being concrete about why. A vision-language model answering a question produces a string that can be compared to a reference. A GUI agent produces a sequence of taps, and success depends on the environment staying still, the action space matching, the episode budget, and the judge deciding whether the task was completed.

UI-Venus's card is more transparent about this than most: it says task completion is judged on task-relevant visual keypoints rather than a coarse look at the final screen, and that training used human–discriminator collaborative verification to filter invalid interactions. That is a description of a harness as much as a model. LingSante's card gives the score and the comparator list, and says GUI evaluations use real apps with changing interfaces and login states. Both are honest. Neither is enough to reconstruct the run.

A team choosing an agent should therefore treat 86.84 and 80.2 as evidence that both models can operate an Android device, not as a ranking. The number that would settle it is a head-to-head on one harness with both checkpoints, and nobody has published one.

What UI-Venus measures that air-VL doesn't

The specialist's advantage is in the rows the medical card omits, and those rows are where computer-use work actually gets decided.

• Grounding — ScreenSpot-Pro 73.0, VenusBench-GD 77.1, OSWorld-G-R 78.5, UI-Vision 53.2. These measure whether the model can point at the right control, which is the failure mode that ruins otherwise competent agents.

• Desktop and web — OSWorld-Verified 70.8, DeskCraft 48.0, WebVoyager 90.8, Online-Mind2Web 74.0, REAL 76.9, Odysseys 77.3.

• Adversarial interfaces — MCA-Bench 75.7, NextGen-CAPTCHAs 47.6, OpenCaptcha 50.7, Spatial-CAPTCHA-Bench 42.8, and an attack success rate of 11.3% on OSHarm and 48.8% on OSBlind, against 25.3% and 79.4% for the Qwen3.5-9B base it was trained from.

None of those appear on LingSante's table. The medical model reports BrowseComp-VL 62.80, MMSearch 83.00 and AndroidWorld 86.84 as its agent-adjacent rows, which is a thin profile for computer use. If your task is filling a portal, navigating a claims system or operating a legacy desktop application, UI-Venus has the finer-grained evidence and a ninth of the memory footprint. LingSante has one strong number and a capability paragraph.

Twenty-five times the memory

The deployment arithmetic is not close. UI-Venus-2-9B is a 9B dense model in four shards that comes up under stock vLLM with a reasoning parser, which is the least dramatic serving story in this family. LingSante needs 249.7 GB resident, tensor parallelism across four 141 GB GPUs, and a YaRN override to reach its advertised context. The ratio in stored weights is roughly twenty-five to one once you count the medical model's full checkpoint.

That cost buys a different capability, not a better version of the same one. LingSante reads imaging studies, drafts Findings and Impression, recognises more than 120 skin conditions, breaks down a meal into ingredients and nutrients, and links blood counts to inflammatory markers. UI-Venus taps buttons. Comparing their price per call is meaningless because they are not substitutes.

Where the two do meet is the medical GUI workflow. LingSante's card describes booking an appointment — choosing hospital, department, doctor and time slot — and buying medication, including prescription entry, address entry and pop-up handling. A team building that product could plausibly use UI-Venus as the interface layer and a separate medical model for the content, which is cheaper and easier to debug than asking one 124B checkpoint to do both.

Quale chiamare

For an agent product, UI-Venus-2-9B is the defensible default: small, 9B dense, stock vLLM, published grounding and safety rows, and a training description specific enough to audit. For clinical content, LingSante-3.1-air-VL is the only one of the two with a claim on the task, provided you have the hardware and read its research-only notice as a constraint.

The composition case is stronger than either alone. A router that holds both behind a single key — one API for 200-plus models — lets an interface agent hand a screenshot or a document to a medical reasoner only when the task needs one, with the routing DSL deciding which model gets the call. Neither model is in the OrcaRouter catalogue today, so this is a statement about how the pipeline should be shaped rather than a hosting claim. The point stands either way: the agent and the reasoner have different budgets and different failure modes, and forcing them into one checkpoint is how you end up paying 250 GB for a button-pusher.

What to watch is whether inclusionAI publishes the harness behind LingSante's AndroidWorld figure. If it does, the 86.84 becomes a comparable result and this matchup gets a real answer. Until then, the honest summary is that the medical model won a row it was not built for, on a benchmark that does not travel, against a specialist that still owns every row that does.

A generated two-column comparison scoreboard titled 'LingSante-3.1-air-VL vs UI-Venus-2-9B — the scoreboard'. The left column, LingSante-3.1-air-VL, reads 'Parameters: 124B total / 5.5B active', 'Weights: 249.7 GB', 'Context: 262,144 tokens', 'AndroidWorld: 86.84', 'Grounding rows: not reported' and 'Licence: MIT'. The right column, UI-Venus-2-9B, reads 'Parameters: 9B dense', 'Weights: four shards', 'Context: 262,144 tokens', 'AndroidWorld: 80.2', 'Grounding rows: ScreenSpot-Pro 73.0' and 'Licence: not declared'. A footer reads 'Vendor-reported on different harnesses; 23 November 2026.' The OrcaRouter logo is composited in a white strip along the bottom edge.The vendor's benchmark-comparison image from the LingSante-3.1-air-VL model card, captured 23 November 2026: a 26-row table in eight columns. The Multimodal Agents block at the bottom reads BrowseComp-VL 62.80, MMSearch 83.00 and AndroidWorld 86.84 for LingSante-3.1-air-VL, with the AndroidWorld row showing Kimi-K3.1-Pro at 86.20, GPT-6.1 Sol at 92.67, Claude Opus 5.5 at 90.09 and Gemini 3.8 Flash at 75.86. Above it, the medical rows include OmniMedVQA 93.64, CheXpert Plus 46.21 and HealthBench Professional 49.59; footers note thinking mode at temperature 0.6 and that comparison models use their highest available reasoning effort.