{"slug": "thesys-ships-oui-1-so-developers-can-run-generative-ui-locally", "title": "Thesys ships OUI-1 so developers can run generative UI locally", "summary": "Thesys released OUI-1, an open-weight generative UI model fine-tuned from Google's DiffusionGemma with 26 billion total and 4 billion active parameters, scoring 71.7% on Thesys' Generative UI Benchmark versus 13.0% for the base model. The model generates interfaces in OpenUI Lang and is designed to run locally, but its benchmark tests structural validity rather than usability.", "body_md": "# Thesys ships OUI-1 so developers can run generative UI locally\n\n**The open-weight DiffusionGemma fine-tune uses 4B active parameters, while Thesys' own benchmark leaves usability untested.**\n\n        By [RuntimeWire Staff](/author/runtimewire-staff)\n        · Published \n\nPrimary source: [OpenUI](https://www.openui.com/blog/oui-1)\n\n## Why it matters\n\nThesys' benchmark measures whether generated interfaces are structurally valid. Adoption will depend on whether those interfaces are usable and help people complete tasks.\n\n[Thesys](https://www.thesys.dev/company?ref=runtimewire), the generative UI startup founded by [Rabi Shanker Guha](https://x.com/rabi_guha?ref=runtimewire) and [Parikshit Deshmukh](https://in.linkedin.com/in/parikshitd?ref=runtimewire), said in its [September 8 launch post](https://www.openui.com/blog/oui-1?ref=runtimewire) that it released OUI-1, an open-weight model built to generate software interfaces rather than another stream of prose.\n\nThe [model](https://huggingface.co/thesysdev/OUI-1?ref=runtimewire) is a fine-tune of Google's DiffusionGemma with 26 billion total parameters and 4 billion active parameters. It writes screens in [OpenUI Lang](https://www.openui.com/docs/openui-lang/comparison?ref=runtimewire), the compact, declarative language behind Thesys' [OpenUI framework](https://www.openui.com/?ref=runtimewire), and can assemble charts, forms, tables, cards and dashboards from a developer's approved component library.\n\nThe weights are distributed under Google's Gemma terms.\n\nGuha previously worked at Google and led engineering at DevRev India. A [Stackforce profile](https://www.stackforce.co/talent/rabi-shanker-guha-ceo-692c807204783d309337b2ce?ref=runtimewire) also says his earlier fintech startup, Payzie, was acquired by Google. Deshmukh previously led design at DevRev and [Recko, the fintech infrastructure provider acquired by Stripe](https://stripe.com/en-ca/newsroom/news/recko?ref=runtimewire).\n\nThat engineering and design pairing is visible in OUI-1. Guha and Deshmukh are trying to make generated interfaces fast enough to feel like ordinary software and constrained enough that developers can safely put them in front of users.\n\n### A model trained against a parser\n\nOUI-1's technical story starts with a failed fine-tune. According to Thesys' [launch post](https://www.openui.com/blog/oui-1?ref=runtimewire), supervised fine-tuning reduced training loss, yet produced longer programs that frequently failed to parse.\n\nThesys then used the OpenUI Lang parser as a verifier. The model generated programs, valid examples were retained, near-misses were repaired according to parser feedback, and the accepted programs became the training set for the next pass. Thesys describes the process as rejection-sampled self-training with repair.\n\nThesys repeated the process across 27 component libraries to produce OUI-1.\n\nThis is a narrower approach than asking a general-purpose model to produce arbitrary React code. OpenUI Lang limits the model to components and properties declared by the developer, while the renderer progressively displays valid output as it arrives. The generated interface cannot introduce arbitrary executable code through the language itself.\n\n### The benchmark tests structure, not product quality\n\nOUI-1 scored 71.7% on Thesys' [Generative UI Benchmark](https://github.com/thesysdev/Generative-ui-bench?ref=runtimewire), compared with 13.0% for the untuned DiffusionGemma base model. The test contains 46 interface briefs with four generations per brief and grades whether the resulting component graph parses, resolves its references, produces a root and uses valid required properties.\n\nOUI-1 also produced 55 valid outputs from 60 previously unseen prompts using a separate AppLess component library, according to [Thesys' launch post](https://www.openui.com/blog/oui-1?ref=runtimewire). The base model produced 23.\n\nThose results support a focused claim: fine-tuning substantially improved DiffusionGemma's ability to produce structurally valid OpenUI programs. Thesys' [benchmark documentation](https://www.openui.com/benchmarks/language?ref=runtimewire) explicitly describes the score as a structural-validity measure rather than a general intelligence test.\n\nThe [benchmark page](https://www.openui.com/benchmarks/language?ref=runtimewire) lists [Qwen3.8-27B](/models/qwen/qwen3.8-27b) at 78.8%, above OUI-1's company-reported 71.7%. The benchmark establishes structural validity, not visual quality, usability or task completion. OUI-1's practical distinction is that Thesys says it can run at FP8 precision on an Nvidia RTX 5090.\n\nThesys calls OUI-1 the world's first model for generative UI. The available evidence does not establish that broad superlative. Google released [A2UI](https://developers.googleblog.com/introducing-a2ui-an-open-project-for-agent-driven-interfaces/?ref=runtimewire) as an open format for agent-generated interfaces in December 2025, and Vercel's [json-render](https://github.com/vercel-labs/json-render?ref=runtimewire) already lets models generate streamed interfaces from constrained component catalogs. Those projects center on protocols and frameworks rather than a dedicated open-weight model, which is the narrower distinction Thesys can substantiate.\n\n### Thesys moves deeper into the stack\n\nBefore OUI-1, Thesys built its C1 generative UI API and opened the underlying language and framework through the [OpenUI repository](https://github.com/thesysdev/openui?ref=runtimewire), which had about 8,600 GitHub stars when checked on September 8.\n\nThe model release gives Thesys control over another layer: the system generating the interface description. OpenUI remains compatible with outside models, but OUI-1 lets developers self-host a model tuned for the framework's language and validation rules.\n\nGuha and Deshmukh financed the initial version of that strategy with a [$4 million seed round](https://www.globenewswire.com/news-release/2024/11/07/2976745/0/en/Thesys-Raises-4M-Led-by-Together-Fund-to-Shape-the-Future-of-AI-Driven-Interfaces.html?ref=runtimewire) announced in November 2024. [Together Fund](https://www.together.fund/portfolio/thesys?ref=runtimewire) led the financing, with participation from 8VC.\n\nThe market around them has since filled out. Google's A2UI focuses on moving generated interfaces across applications and trust boundaries. Vercel's json-render supports constrained catalogs across web, mobile, email, PDF and other targets. CopilotKit raised a [$27 million Series A](https://www.copilotkit.ai/blog/series-a?ref=runtimewire) in May 2026 to expand its agent-to-frontend stack and AG-UI protocol.\n\nOUI-1 gives Thesys a clear position inside that group: an open-weight, specialized model paired with a compact interface language and streaming renderer. Its next test will happen outside the parser, when developers put generated screens in front of users who care less about valid component graphs than whether the software helps them finish the job.", "url": "https://wpnews.pro/news/thesys-ships-oui-1-so-developers-can-run-generative-ui-locally", "canonical_source": "https://runtimewire.com/article/thesys-ships-oui-1-local-generative-ui-model", "published_at": "2026-09-08 20:07:03+00:00", "updated_at": "2026-09-08 20:21:42.197515+00:00", "lang": "en", "topics": ["artificial-intelligence", "generative-ai", "ai-products", "ai-research"], "entities": ["Thesys", "OUI-1", "DiffusionGemma", "Google", "Rabi Shanker Guha", "Parikshit Deshmukh", "OpenUI Lang", "OpenUI framework"], "alternates": {"html": "https://wpnews.pro/news/thesys-ships-oui-1-so-developers-can-run-generative-ui-locally", "markdown": "https://wpnews.pro/news/thesys-ships-oui-1-so-developers-can-run-generative-ui-locally.md", "text": "https://wpnews.pro/news/thesys-ships-oui-1-so-developers-can-run-generative-ui-locally.txt", "jsonld": "https://wpnews.pro/news/thesys-ships-oui-1-so-developers-can-run-generative-ui-locally.jsonld"}}