I made WEBGEN-OSS-20B, a model that generates clean websites from your prompts (huggingface.co)

Developer Tesslate released WEBGEN-OSS-20B, a 20-billion-parameter open-source causal language model tuned specifically to generate clean, production-lean single-file websites. Unlike general-purpose code LMs, it prioritizes semantic HTML, sensible spacing, and modern component blocks (hero, grids, pricing, FAQ) and defaults to Tailwind utility classes. The project includes a ready-to-run Hugging Face/Transformers example (model_id "Tesslate/WEBGEN-OSS-20B") showing how to load the model with AutoTokenizer/AutoModelForCausalLM, run inference in bfloat16 for lower memory, and use device_map="auto" for multi-GPU or mixed-device deployment. Significance: by constraining scope to HTML/CSS/Tailwind and optimizing for structural consistency, WEBGEN-OSS-20B is small enough to run locally for fast iteration while producing production-usable static pages (the example enforces "no external JS" and generates a single-file landing page). Typical generation knobs in the demo—temperature 0.7, top_p 0.9, max_new_tokens 2000—show it can produce lengthy, opinionated layouts. This makes it useful for rapid prototyping, scaffolding static sites, or integrating into dev tools, though outputs still need human review for accessibility, security, and correctness.