Generative interfaces put to the test of specialisation: Thesys's OUI-1 model

While open-weight models keep evolving mainly along the axis of parameter count, Thesys has tried the route of vertical specialisation, announcing OUI-1 and presenting it as the first open-weight model dedicated to Generative UI — a claim that depends on how you draw the boundaries of the category, and that nobody has verified independently. Built through a LoRA finetune of Google's DiffusionGemma 26B-A4B-it, whose weights were then merged into the bf16 safetensors format, the model counts 26 billion total parameters, 4 billion of them active, and a context window of 16,384 tokens. The goal stated in the Hugging Face model card is to make it possible to generate interfaces for applications built on the OpenUI framework locally, reducing the dependence on remote APIs. On the distribution side, the weights remain subject to Google's Gemma Terms of Use, a proprietary licence with usage constraints clearly distinct from the OSI-approved open source licences — unlike the OpenUI framework itself, released under MIT.

The performance figures released by the company credit OUI-1 with 71.7% on the Generative UI Benchmark, against 13.0% for the base model. The evaluation rests on openui-lang, a declarative language that, by Thesys's own measurements, cuts the tokens required by up to 67% compared with JSON. The same comparison table, however, shows a generalist model such as Qwen3.8 27B scoring higher, at 78.8%: what OUI-1 claims is therefore not the absolute score, but getting close to it with 4 billion active parameters. The benchmark itself — 46 briefs spread across five complexity tiers — also has a narrow perimeter: the metric measures only the formal correctness of the output, things like the absence of orphan components or parsing errors, without capturing the usability or the aesthetic quality of the design produced.

On the execution side the numbers do not entirely add up. The requirements list roughly 25.8 GiB of VRAM to run in FP8 quantisation through vLLM, which according to Thesys brings a consumer GPU such as the RTX 5090 within reach; the original tests, though, were run on a single A100. The latency figures do not line up either: the announcement post reports 1.9 seconds per output, while the model card states about one second per screen on a single GPU — the two measurements cannot be traced back to the same configuration. In the technical discussion that emerged on Hacker News, doubts were raised about the consistency of interfaces regenerated at every session, about the debugging difficulties tied to non-deterministic output, and about the absence of any visual evidence in the official announcement.

Moving Generative UI from remote APIs to a compact, self-hostable model is a rational architectural choice for keeping compute costs down. It remains the case that when a single party defines the model, the output syntax and the evaluation benchmark, the resulting metric measures above all adherence to a perimeter of rules it built for itself. — Olya