Molmo2-4B vs UI-Venus-2-9B: One of Those Names Is a Size, and It Is Wrong

The model is called UI-Venus 2 at 9B. It is not 2.9B, and it is not a 9B version of a 2.9B model. The string in the filename is a generation number glued to a parameter count — inclusionAI/UI-Venus-2-9B — and it has been read backwards by enough radar feeds and listing pages that a comparison against Molmo2-4B now looks like two small models in the same weight class. It is not. UI-Venus-2-9B is a 9B GUI agent initialised from Qwen3.5-9B, roughly twice the size of Molmo2-4B's 4.85B, and the two of them do different jobs entirely. If you actually want the small GUI model from that family, it is UI-Venus-1.5-2B, which is where the 2B number belongs. Two specialist models with different output contracts are a routing problem. no markup over the provider's list price Failover covers the case where one of them is having a bad afternoon.

Both models are older than the framing suggests. Molmo2-4B was announced by the Allen Institute for AI on 11 December 2025 and its weights went up on 14 December. UI-Venus-2-9B's weights landed on 26 August 2026 alongside a technical report, arXiv 2609.00028. What changed recently on the Molmo side is tooling rather than weights: native Transformers support merged on 9 October 2026 and shipped in v5.20.0 on 15 October, and a robotics fine-tune, Molmo2-4B-tb, appeared on 5 November. Neither is a launch, and neither turns Molmo2-4B into a GUI model.

What UI-Venus 2 at 9B actually is

It is a closed-loop interface agent. The input is a screenshot; the output is a reasoned action — tap here, type this, scroll — and the loop continues until the task is done. Training runs in three stages over five task families (grounding, CAPTCHA, mobile, web and computer), ending in multi-teacher on-policy distillation that folds domain specialists into one policy. The environment coverage is the substance of the release: 170+ mobile apps across Chinese and English, a web pool of 4,000+ domains in 19 categories, and desktop OS control built from scratch.

The vendor's own numbers, all from the model card and its technical report:

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

• Mobile agents — AndroidWorld 80.2, MobileWorld 65.8 at standard steps and 75.2 at a 100-step budget, MobileGym 52.7, VenusBench-Mobile 46.5, KnowUBench 56.5, MemGUI 62.6.

• Desktop — OSWorld-Verified 70.8, DeskCraft 48.0, and OSWorld 2.0 at 0.0 binary accuracy with a 7.5 partial score over 108 tasks under a 150-step budget. That zero is worth reading carefully; it is a hard binary metric, not a typo.

• Web — WebVoyager 90.8, Online-Mind2Web 74.0, REAL 76.9, Odysseys 77.3 average with 62.0 perfect.

• Safety — attack success rate 11.3% on OSHarm and 48.8% on OSBlind, against 25.3% and 79.4% for the Qwen3.5-9B base.

One independent check exists. The PointerBench diagnostic benchmark published in October 2026 scores UI-Venus-2-9B at Sheets 64.2, Text 25.8 and Pro 73.6, a 54.5 average, and describes it as the strongest native GUI-grounding specialist while flagging text-level grounding as its weakest area. That is a third-party result on a different task mix, and it is a useful corrective: the ScreenSpot-Pro headline is not the whole picture.

What Molmo2-4B actually is

A general-purpose vision-language model with a point output. 4,850,869,200 parameters, a SigLIP 2 so400m vision tower over Qwen3-4B-Instruct-2507, 36,864 positions, 19.42 GB of float32 weights across four shards. It handles images, multiple images and video, and it returns coordinates scaled to 0–1000 of image width and height, with timestamps when the input is a video. Its vendor aggregate is 62.8 across 15 academic benchmarks.

What it does not have is a published GUI number. Molmo2-4B does not appear in the UI-Venus technical report's tables, in the PointerBench evaluation, or with a ScreenSpot-Pro result on its own card. The closest Molmo-family GUI figure belongs to a different model: MolmoPoint-GUI-8B, an 8B GUI-specialised fine-tune in the March 2026 MolmoPoint line, reports 61.1 on ScreenSpot-Pro. Quoting that as a Molmo2-4B result would be a category error, and it is the single most likely error for a reader to make when comparing these two.

The overlap is smaller than the naming suggests

• Size — 4.85B dense for Molmo2-4B against a 9B dense model initialised from Qwen3.5-9B.

• Inputs — image, multi-image and video for Molmo2-4B; a screenshot for UI-Venus-2-9B, in a loop, with the previous action fed back.

• Outputs — sparse points in normalised coordinates for Molmo2-4B; a structured action for UI-Venus-2-9B, parsed by a harness that then drives the device.

• Strength — natural-image and video grounding, counting and tracking for Molmo2-4B; element localisation and multi-step navigation on high-resolution interfaces for UI-Venus-2-9B.

• Published GUI grounding — none for Molmo2-4B; ScreenSpot-Pro 73.0 for UI-Venus-2-9B.

• Licence — Apache 2.0 weights with a third-party data caveat for Molmo2-4B; the UI-Venus-2-9B model card publishes no licence field at all, which is a gap to resolve before anything commercial.

That last row is not a detail. A 9B model derived from Qwen3.5-9B with no licence field on its model card is not something a legal review can sign off on the strength of a README. The card's own metadata does not help either: its Safetensors badge reads "1.47M params", which is not a plausible size for a checkpoint this large and looks like a shard-metadata artifact. Ask the maintainers about both before you plan around it.

If you want GUI grounding from the Molmo line

The honest path is not Molmo2-4B. Ai2's own GUI work sits in the MolmoPoint family, where MolmoPoint-GUI-8B reports 61.1 on ScreenSpot-Pro against UI-Venus-2-9B's 73.0, and the two numbers at least come from the same benchmark. Even then, the comparison is cross-harness: different action scaffolds, different screenshot preprocessing, and in UI-Venus's case a documented assumption that reasoning is disabled and temperature set to zero for direct point localisation. A 12-point gap on a benchmark that sensitive to harness choices is a signal, not a verdict.

What Molmo2-4B is actually good for in an agent stack is the perception step that is not a button. Finding the object a user described in a camera frame, pointing at it across a video, counting instances, tracking something through time — those are the outputs the model was trained for, and none of them is a UI element on a 4K screenshot. Teams often end up running both kinds of model: a specialist for the interface, a generalist for the scene.

That is a routing problem more than a modelling one. The two models have incompatible output contracts and different hardware profiles, and running them behind a single key across both — with failover when one endpoint degrades and no markup over the provider's list price — removes the part of the stack that has nothing to do with vision. It is also worth saying plainly that neither of these models is in the OrcaRouter catalogue; both are weights you host yourself, and the routing layer sits around them rather than under them.

The short version

If the task is "operate this app," UI-Venus-2-9B is the specialist and Molmo2-4B has nothing to offer on that benchmark. If the task is "point at the thing in this photograph or this video," UI-Venus-2-9B has nothing to offer and Molmo2-4B is the reason the comparison exists. And if what you actually wanted was a small GUI model to run cheaply, the correct comparison was never this one — it is UI-Venus-1.5-2B at 2B, against whatever compact generalist you can fit beside it.