InternLumina-U2 vs Tencent UI-Mate-27B: The Desktop Agent You Can Run Now vs the Unified Vision Model You Can Only Read

Tencent UI-Mate-27B and InternLumina-U2 are two quiet Apache-2.0 drops from Chinese labs, published a fortnight apart, and they are not competitors — which is precisely why they are worth comparing. Tencent UI-Mate-27B, released as open weights on 14 August 2026, is a desktop agent: it watches the screen and emits mouse and keyboard actions. InternLumina-U2, published as inference code on 1 September 2026 by Shanghai AI Laboratory, is a would-be unified vision model: it claims to read images, video and 3D and to generate and edit images. One acts on what it sees; the other, when its weights finally exist, will talk and draw about what it sees. If your job is operating a desktop, only one of these two can do it today — and it is not the one with the fancier architecture.

The agent that acts: Tencent UI-Mate-27B

UI-Mate-27B is a 27-billion-parameter open-weight foundation GUI agent from Tencent's HY Frontier multimodal agent team, fine-tuned from Qwen3.6-27B. It observes live screenshots, reasons over the current screen state, and outputs structured keyboard and mouse actions in a normalized coordinate space — the artifact of a model trained to operate a real desktop, not to chat about one. Its distinguishing feature is demonstration-guided execution: show it a workflow once and it adapts the recorded demonstration to the task at hand, treating the live screenshot as authoritative when the interface differs from what was recorded. Tencent's reported headline scores are OSWorld-Verified 77.0 and WindowsAgentArena 66.2 — figures that would top those 2026 leaderboards if independently reproduced — alongside a spread of OSWorkerBench numbers. The weights are real and downloadable: twelve safetensors shards on Hugging Face, self-hostable through vLLM or SGLang on a couple of GPUs, and the model card carries an important caveat that it is an agent checkpoint rather than a standalone visual-chat model — point it at "describe this image" and you are using it wrong.

The eyes that are promised: InternLumina-U2

InternLumina-U2 is a different species entirely. It is a 16B-parameter sparse mixture-of-experts diffusion model (1B active) that the lab intends as one model for omni-visual understanding, image generation and image editing — a fully-discrete eight-codebook design where a single backbone both reads an image, chart, video or 3D asset and writes new pixels. If your mental model is a GUI agent, InternLumina-U2 is the opposite pole: it does not want to click anything, it wants to understand everything and draw on request. Its published form on 1 September 2026 is inference code and an architecture — six task entry points in a GitHub repository, a project page with a preliminary benchmark table, an empty Hugging Face stub — and its weights, training code and technical report are all still marked coming soon. Nobody can run it. The gap between the two models is not quality; it is existence.

The scoreboard

UI-Mate-27B's figures are Tencent-reported and unreproduced; InternLumina-U2's are lab-reported, preliminary and partial. Every row carries its provenance.

• Job — Tencent UI-Mate-27B: operate a desktop — read live screenshots, emit mouse and keyboard actions. InternLumina-U2: perceive and create — understand images/video/3D, generate and edit images.

• Weights — Tencent UI-Mate-27B: public since 14 August 2026, twelve safetensors shards, Apache-2.0. InternLumina-U2: not public — empty Hugging Face repo, weights "coming soon."

• Scale — 27B dense, fine-tuned from Qwen3.6-27B vs 16B total / 1B active sparse MoE diffusion model.

• Output — Tencent UI-Mate-27B: structured pyautogui-compatible actions in a normalized coordinate space. InternLumina-U2: claims text, text-to-image up to 1024×1024, and instruction-based image editing.

• Headline numbers — Tencent UI-Mate-27B: OSWorld-Verified 77.0, WindowsAgentArena 66.2, OSWorkerBench gains from demonstrations (all Tencent-reported). InternLumina-U2: ChartQA 86.52, GenEval 0.81, MathVision 33.22 (lab-reported, preliminary).

• Provenance caveat — Tencent UI-Mate-27B: the card itself warns it is an agent checkpoint, not a standalone visual-chat model. InternLumina-U2: the project page labels its own results "preliminary, partial."

• Serving reality — Tencent UI-Mate-27B: self-host via vLLM or SGLang on ~2 GPUs; no hosted inference provider currently deploys it. InternLumina-U2: not servable by anyone — the released driver expects checkpoints that are not in the repository.

Two different flavors of unproven

Both models are unproven, and the word does not mean the same thing twice. Tencent UI-Mate-27B is unproven in the ordinary sense of a new model: its headline scores are the vendor's own, nobody has independently rerun them, and its download count is tiny next to the claims — the usual posture for a four-day-old-at-first-glance checkpoint that has since grown a modest community. But it is unproven in the testable way. Because the weights exist, the 66.2 and the 77.0 can be checked this week by anyone with two GPUs and a Windows test environment, and the demonstration-guided claims can be reproduced or refuted on real workflows. InternLumina-U2 is unproven in a stricter sense: it is untestable. Its architecture is public and readable, its numbers are not — there is no artifact that could confirm them, and the lab's own partial table already shows it trailing a named rival on at least one generation benchmark. A vendor number attached to a downloadable file is a claim awaiting a referee. A vendor number attached to a roadmap is a hope.

The tencent/UI-Mate-27B Hugging Face card, captured 27 August 2026 — the Qwen3.6-27B base, the vendor-reported OSWorld and WindowsAgentArena scores, and the note that no inference provider currently deploys it.

The natural division of labor: an actor and its eyes

Set side by side, the two models sketch the shape of a reliable automation pipeline. The hard problem with GUI agents is not the average case; it is the agent quietly doing the wrong thing on an unexpected screen state and no one noticing. That is exactly the job a cheap, trustworthy vision model exists for — verify the screen state before and after every action, confirm the dialog that appeared is the dialog the agent thinks appeared, and halt the loop when reality and the agent's model of reality diverge. Tencent UI-Mate-27B is the actor; a unified vision model of the kind InternLumina-U2 claims to be is the natural verifier, able to read the same screenshots the agent acts on. When — and only when — InternLumina-U2's weights actually land and its understanding claims survive independent testing, that is the role it could fill alongside an agent rather than against it. The two are not rivals for a job; they are the two halves of a job.

The InternLM/InternLumina-U2 GitHub repository, captured 2 September 2026 — the repo landing page showing the per-task inference scripts, including the image-understanding entry point a screen-state verifier would be built from; the weights that would make it runnable are not in the tree.

What a routing layer does for an agent-plus-eyes stack

Neither model is hosted on OrcaRouter, and we will not imply otherwise — Tencent UI-Mate-27B has no hosted provider anywhere, and InternLumina-U2 has no weights for any provider to host. The routing pattern that applies is the one for exactly this architecture: an agent that acts and a vision model that verifies, composed so the expensive or risky step is gated on the cheap reliable one. The routing DSL expresses that as a single logical call — act, then verify, then proceed or halt — instead of bespoke glue between two model providers, and automatic failover covers the realistic failure mode: a GUI agent like UI-Mate-27B is exactly the kind of unproven checkpoint you audition on a test path first, with the call falling through to a proven model the moment the new one misbehaves. One API across 200+ hosted models, provider list price passed through at 0% markup, and the unproven pieces of the stack kept behind safety rails while they earn your trust.

The bottom line

If you are building something that operates a desktop, Tencent UI-Mate-27B is the only one of these two that exists in a usable sense — runnable today on your own GPUs, with impressive but unreproduced scores that you can actually go test. If you are building something that needs a model to understand and draw about what it sees, InternLumina-U2 is a promising architecture waiting on its weights — worth reading now, worth nothing as a dependency until the safetensors appear. Watch the two for the day the division of labor becomes possible: an actor on your desktop, a set of verified eyes watching it, and a routing layer keeping them honest.