MiniCPM-V 4.7 vs UI-Venus 2.9B: A General Vision Model Against a Purpose-Built GUI Agent

MiniCPM-V 4.7 and UI-Venus 2.9B are both vision-language models that look at a screenshot and produce text, and there the resemblance ends — because one of them was built for exactly that job. UI-Venus 2.9B is a general-purpose GUI agent from inclusionAI, released August 26, 2026, whose entire design centres on observing an interface, reasoning about task state, emitting an action, and incorporating the resulting feedback; it ships with a technical report, benchmark tables across GUI grounding, mobile, web, computer-use and CAPTCHA tasks, and an explicit evaluation of how often it can be talked into a dangerous action. MiniCPM-V 4.7 is a 35.2-billion-parameter sparse mixture-of-experts checkpoint uploaded by OpenBMB on October 6, 2026 with no card, no licence, and no benchmark. Comparing a specialist against an unmeasured generalist is a somewhat unusual exercise, and this page is explicit about where the comparison holds and where it does not.

Two very different kinds of release

UI-Venus 2.9B is inclusionAI/UI-Venus-2-9B, a 9B-parameter model initialised from Qwen3.5-9B and post-trained specifically as a GUI agent. The card describes a closed loop — observe the interface, reason about the task state, execute an action, fold the feedback into the next decision — trained across three stages: multimodal mid-training on large-scale simulated mobile, web and desktop trajectories; offline reinforcement learning split across five task families; and multi-teacher on-policy distillation that consolidates domain-specialised teachers into one policy. It is evaluated on GUI grounding, mobile, web, computer-use and CAPTCHA benchmarks, and separately on consequential-action safety: the card reports an attack success rate of 11.3% on OSHarm and 48.8% on OSBlind against 25.3% and 79.4% for its Qwen3.5-9B base. OpenBMB's own sibling, incidentally, looked at similar ground: the MiniCPM organisation has an AgentCPM-GUI project from January 2026, so this is a lane both labs have entered.

MiniCPM-V 4.7 is openbmb/MiniCPM-V-4.7-35B-A3B, 35,212,875,824 BF16 parameters across 70.4 GB and sixteen shards, uploaded October 6, 2026.

Sparse MoE text backbone tagged qwen3_5_moe_text — 256 experts, 8 per token — over 40 layers with a three-linear-to-one-full attention pattern, a 27-layer in-house vision tower with 16× downsampling and up to nine image slices, and a 256K context window. No README, therefore no licence and no benchmarks. Three likes, zero downloads, empty discussions.

The comparison, with the gaps left open

• Purpose — UI-Venus 2.9B: a GUI agent, trained end-to-end for interface interaction. MiniCPM-V 4.7: a general vision-language model; what it is optimised for is not stated.

• Parameters — UI-Venus 2.9B: 9.41B, dense. MiniCPM-V 4.7: 35.2B total, sparse, 8 of 256 experts per token.

• Base model — UI-Venus 2.9B: initialised from Qwen3.5-9B, a dense Qwen text stack. MiniCPM-V 4.7: a Qwen3.5-derived MoE text stack, class qwen3_5_moe_text. Different branches of the same family tree.

• Interface grounding — UI-Venus 2.9B: the core competency, with dedicated grounding and CAPTCHA task families in training and a keypoint-grounded verification system using multi-model voting. MiniCPM-V 4.7: no grounding-specific claim, and no coordinate-output format documented.

• Safety evaluation — UI-Venus 2.9B: reports ASR of 11.3% on OSHarm and 48.8% on OSBlind. MiniCPM-V 4.7: none.

• Evidence — UI-Venus 2.9B: a technical report on arXiv, a project page, a GitHub repo, and benchmark tables. MiniCPM-V 4.7: a config file.

• Tooling — UI-Venus 2.9B: vLLM quick-start with a reasoning-parser flag, plus GGUF conversions from the community. MiniCPM-V 4.7: nothing yet.

Why a GUI agent is not just "a VLM that looks at screens"

The gap between these two models is not mainly about parameter count, and it is worth spelling out because it is the thing that makes the matchup interesting rather than merely lopsided.

A screenshot task has a property most vision benchmarks do not: the output has to be executable. When UI-Venus 2.9B is asked to tap a button, it has to emit a coordinate or a structured action that the environment can actually apply, and a small error in grounding is not a wrong answer on a leaderboard, it is a failed task and a corrupted state. That is why the UI-Venus 2 card spends so much of its length on verification: task completion is judged on task-relevant visual keypoints rather than a holistic glance at the final screen, and heterogeneous judges vote to reduce single-judge bias. The point of that machinery is to make the reinforcement-learning reward signal robust to reward hacking — a model that learns to satisfy a lazy verifier rather than to complete the task.

It also explains the safety section.

An agent that can operate a phone or a desktop has an attack surface that a captioning model does not, and reporting an attack success rate is the minimum a lab should do when shipping one. UI-Venus 2.9B reports 11.3% on OSHarm and 48.8% on OSBlind — the second figure is high in absolute terms, and the card's framing is a comparison against its base model rather than an absolute claim of robustness. Read it as "meaningfully better than where it started," not as "safe."

MiniCPM-V 4.7's architecture has nothing to say about any of this. Linear attention over a long context would help an agent that has to remember a long interaction history, and the sparse MoE would help throughput if you are running many episodes. But an architecture that would be useful for an agent is not an agent, and no part of the released artifact documents action output, grounding accuracy, or safety behaviour.

The honest assessment of the MiniCPM side

It is tempting to fill this section with 4.6's numbers. Resist it. MiniCPM-V 4.6 is a 1.3B dense model on a Qwen3.5-0.8B backbone with a switchable 4x/16x visual compression scheme, and its advertised results — a score of 13 on the Artificial Analysis Intelligence Index, vendor-reported, at a fraction of the token cost of comparable small models — describe that model. MiniCPM-V 4.7 changes the backbone, changes the attention pattern, and changes the default visual compression. None of 4.6's measurements transfer, and none of them describe a GUI agent in the first place.

What can honestly be said about MiniCPM-V 4.7 is narrow and factual: the weights exist, the shape is unusual for the family, the context window is long, and the release is incomplete. The absence of a licence is the practical blocker; the absence of benchmarks is the evaluative one. And there is one modest reason for interest rather than indifference — OpenBMB builds GUI tooling, AgentCPM-GUI among it, so a long-context sparse MoE vision model from that lab is not implausible as a future agent backbone. That is a hypothesis about intent, and the repository does not confirm it.

What you would pick, and when

For anything involving a screenshot and an action — tapping, typing, scrolling, navigating an app or a website, extracting data from a UI — UI-Venus 2.9B is the only one of the two that addresses the task. It has the training, the verification machinery, the reported safety numbers, and enough tooling to stand it up on vLLM today. Nothing about MiniCPM-V 4.7 suggests it is competing there, and without a card you cannot even check whether it emits coordinates.

For general multimodal work — describing images, reading documents, long-context analysis over image-heavy material — the comparison is not decidable yet, because one side has no published quality evidence. If you need that today, UI-Venus 2.9B is at least measurable, though it was tuned for interfaces rather than for document reasoning and is not an obvious first choice for that job either.

The pattern in both cases is the same: a self-hosted model that handles the routine work, and a hosted frontier model for the hard cases it hands off. That handoff is where a router earns its place — one key across 200-plus hosted models, each passed through at the provider's list price with no markup of ours, failing over automatically when a provider degrades. Neither UI-Venus 2.9B nor MiniCPM-V 4.7 is hosted on OrcaRouter; both are weights you run yourself. The router is what your agent talks to when it decides the local model is not enough.

Where this leaves you

This is not a close call, and the reason is structural rather than a judgement about quality. UI-Venus 2.9B is a specialist with a report, a licence, benchmark tables, safety evaluation and a serving recipe. MiniCPM-V 4.7 is a generalist with a configuration file. One of those is a product and the other is a promise, and the only responsible way to compare them is to say so. Watch the repository: if OpenBMB publishes a card that shows interface grounding and action output, then a 256K-context sparse MoE against a distilled 9B agent becomes a genuinely interesting question. Until then, the answer to "which should I use" is the one that exists.