MiniCPM-V 4.7 35B-A3B vs UI-Mate-27B: One Reads the Screen, the Other Presses the Button

The honest first sentence about MiniCPM-V 4.7 35B-A3B against UI-Mate-27B is that they are not competitors. Both were pushed to Hugging Face without an announcement, both carry a permissive licence, both are trained on Qwen-family backbones, and both are described in their own documentation as strong at understanding what is on a screen — which is why they get filed together. But one is a general vision-language model that reads an image and answers in text, and the other is a GUI agent that reads a screenshot and emits pyautogui-compatible mouse and keyboard actions. MiniCPM-V 4.7 35B-A3B can tell you what a dialog box says. UI-Mate-27B can close it. Everything else in this comparison follows from that one difference, and the useful question is not which model wins — it is which half of the job you are missing.

Both are also early. UI-Mate-27B landed on 14 August 2026 with a paper two days behind it and no launch post from the lab; it had zero downloads on Hugging Face when we first covered it, and there is still no hosted API anywhere. MiniCPM-V 4.7 arrived on 6 October with no card at all, and OpenBMB did not publish a model card, a licence tag or a claims section until 12 October. Neither has been independently evaluated. Every performance figure in this piece is labelled with who reported it, and on both sides that is the model's own lab.

The same body, two different jobs

Start with what makes the pairing feel natural. Both are mid-to-large open-weight multimodal models with Apache-2.0 licences. Both sit on the same upstream lineage — UI-Mate-27B is fine-tuned from Qwen3.6-27B, and MiniCPM-V 4.7 35B-A3B's language stack is a Qwen3.5-family sparse MoE. Both are aimed at the same broad application area, screens and documents and application state, and both were released into silence.

Then look at what each one outputs. MiniCPM-V 4.7 returns tokens of text. UI-Mate-27B returns a structured action — a click, a keystroke, a scroll, a wait, a task-complete signal — in a normalised 1000×1000 coordinate space that the reference agent rescales to the real screenshot resolution, so predictions survive display-scaling differences between machines. The model card is explicit that UI-Mate is "an agent checkpoint rather than a standalone visual-chat model" and that the accompanying harness should be used for anything real.

That distinction decides most real deployments. If your pipeline is "read this screenshot and tell me what is wrong with it," a general VLM is the right tool and an agent is overkill with a worse failure mode. If your pipeline is "make the application do the thing," a VLM alone leaves you writing the entire action layer yourself — coordinate mapping, action parsing, error recovery, retry logic — which is the part UI-Mate spent its reinforcement learning budget on.

What MiniCPM-V 4.7 35B-A3B is, as of now

OpenBMB's first MoE in the MiniCPM-V series: 256 experts with 8 active per token across 40 layers, a 262,144-token context window and a 248,144-token vocabulary, or roughly 3B active parameters out of 35.2B total. Vision runs through a SigLIP2 module with LLaVA-UHD v4 intra-ViT early compression and a selectable mixed 4x/16x visual token compression mode — 16x by default for throughput, 4x when fine detail is the point. It understands single images, multi-image sets and video, and returns text.

The card's headline claim is that this is the first open-source MoE model with lossless, smooth on-device inference — Lookahead Expert Scheduling, fine-grained layer-wise quantisation, and speculative sampling through separate DSpark drafts — at a reported 13-plus tokens per second on an iPhone 17 Pro Max. That figure is vendor-reported, unreproduced and unqualified. A second claim, that the model beats Qwen3.6 35B-A3B's overall Instruct-mode score in less inference time, is rendered inside a chart image rather than a table, so it cannot be quoted as a number.

What is concrete is the distribution. Base, FP8, int4, AWQ, GPTQ and GGUF weights all exist; SGLang, vLLM, llama.cpp, Ollama and OpenVINO are supported; iOS, Android and HarmonyOS apps exist; Rockchip, Houmo, Intel, AMD and Arm are listed for day-zero silicon; and the card documents tool calling, with a worked function-calling example in the serving section. That last detail matters for this comparison: MiniCPM-V 4.7 can be the perception and reasoning layer inside an agent you build, even though it will not act on its own.

What UI-Mate-27B is, as of now

the vendor's 27-billion-parameter foundation GUI agent, published on 14 August 2026 alongside a 9B sibling and a demonstration-guided variant, UI-Mate-democua-27B. Training was supervised fine-tuning followed by online reinforcement learning in executable GUI environments — the model learned by actually driving desktops, not from static screenshots. Its action space covers mouse, keyboard, scrolling, waiting, user interaction and completion signals, and the emitted actions are pyautogui-compatible, so the harness can execute them directly.

The differentiator the vendor leans on is demonstration-guided execution: record a workflow once and the agent adapts it to the current screen, with the live screenshot treated as authoritative rather than replaying a fixed script. The supporting material is unusually complete for a silent drop — a project page with a demo video and a macOS Apple Silicon app, a GitHub repository containing the official prompt, response parser and interaction harness, and arXiv preprint 2608.15930 submitted two days after the weights went up.

The headline figures are a vendor-reported WindowsAgentArena score of 66.2 and OSWorld-Verified of 77.0, which would sit at the top of what has been published on that benchmark this year if they hold. They have not been reproduced by anyone outside the lab. And the deployment picture is thin in the way early agent releases usually are: no hosted API, a recommended serving path of vLLM behind an OpenAI-compatible endpoint, and a deliberate instruction to use the harness rather than call the checkpoint raw.

Eight dimensions, both sides

• Output — text, from MiniCPM-V 4.7 35B-A3B; structured pyautogui-compatible actions, from UI-Mate-27B.

• Input — text, images and video, against live screenshots plus an instruction. Neither takes audio.

• Base model — a Qwen3.5-family sparse MoE for the MiniCPM model, Qwen3.6-27B for UI-Mate.

• Size and shape — 35.2B total with 3B active across 256 experts; 27B dense. The sparse one is cheaper per token, the dense one is simpler to serve.

• Context — 262,144 tokens against a default of five retained screenshots, which is the single largest architectural difference between them and the reason one can read a long document set and the other cannot.

• Licence — Apache-2.0 on both, which makes the choice a technical one rather than a legal one for a change.

• Headline claim, both vendor-reported — 13-plus tokens per second on an iPhone 17 Pro Max, against WindowsAgentArena 66.2 and OSWorld-Verified 77.0. Different benchmarks, no shared scale, neither reproduced.

• Available today — six quantisation formats, five runtimes and mobile apps, against zero downloads at first coverage, no hosted API, and a harness that the vendor advises you to use.

Where they actually overlap

There is one job both can attempt: understanding a UI screenshot well enough to answer a question about it. Ask "which field failed validation" and either model can plausibly answer. Ask "fix it" and only UI-Mate-27B produces an action, and only if you have set up the draw — the screenshots loop, the parser, the executor, the coordinate rescaling, and the retry policy when the click lands somewhere unexpected.

The reverse overlap is weaker but real. MiniCPM-V 4.7's tool-calling support means you can build an agent around it — perception and planning on the vision model, with your own action layer bolted on through function calls. That is a legitimate architecture and it gives you total control over the action space. It is also, roughly, reinventing what UI-Mate already spent reinforcement learning on, and doing it without GUI-specific training data. Teams are better off choosing on which layer they want to own than on which model is superficially better at screenshots.

Serving both, and where a router fits

Neither model is hosted anywhere, including here — MiniCPM-V 4.7 35B-A3B and UI-Mate-27B both appear nowhere in OrcaRouter's catalogue, and nothing in this article should be read as a claim that we route them. Both are checkpoints you download and run yourself, and for UI-Mate that means two things: about 27B of weights, and an OpenAI-compatible endpoint, because the vendor's own README recommends serving it with vLLM and the harness expects a compatible base URL.

That endpoint is the interesting part for cost planning, because the harness does not care which model answers as long as the endpoint behaves. OrcaRouter puts 200-plus hosted models behind one key at each provider's list price with no markup of ours, with automatic failover between provider routes, which makes it a drop-in base URL for the planning side of an agent — and the routing DSL lets you compose a cheap model for routine steps and a stronger one for the decisions that matter in one call rather than two integrations. The closest hosted shape to the MiniCPM model's cost profile is Qwen3.6 35B A3B — 35B total, 3B active, 262K context — at $0.25 per million input tokens and $1.48 per million output tokens. Whether that beats self-hosting a 70 GB sparse MoE is a question you can answer in an afternoon, which is not true of either model in the headline.

Which one you want

Take UI-Mate-27B if you need a model that acts — desktop automation, RPA replacement, long-horizon application workflows — and if you accept the trade that comes with it: a vendor-reported benchmark, no independent reproduction, no hosted route, and a harness you must adopt rather than interrogate. The upside is that everything about the release is unusually well documented, from the paper to the prompt to the parser, and the 9B sibling gives you a cheaper model to prototype on.

Take MiniCPM-V 4.7 35B-A3B if you need perception — long documents, screenshot fleets, video — and you are prepared to own the action layer yourself. You get the permissive licence, six quantisation formats, five runtimes, a 262K window and a phone-class deployment story that has not been verified by anyone outside OpenBMB.

What would change this: a reproduced OSWorld or WindowsAgentArena run for UI-Mate, which would turn its central claim from marketing into a measurement; and an evaluation table for MiniCPM-V 4.7, which does not exist at all. Until one of those arrives, the decision belongs to your architecture diagram rather than to either scoreboard — a decoder or a hand, and you should know which one you are short of.