How to use from
Lemonade
Pull the model
# Download Lemonade from https://lemonade-server.ai/
lemonade pull ccore/CCORE-Web-Coder-4B:Q8_0
Run and chat with the model
lemonade run user.CCORE-Web-Coder-4B-Q8_0
List all available models
lemonade list
Quick Links

Important note · Next releases

CCORE Web Coder 4B

CCORE Web Coder 4B turns a brief into a single React HTML document for a website, slide deck, or informational page. CCORE fine-tuned the model from Qwen3.5-4B-Base.

Preview · Approximately 26% of the projected training tokens. Available as ccore-web-coder-Q8_0.gguf (Q8_0, 4.48 GB).

What you build on your computer stays on your computer.

After Dark, a rooftop observatory website Fieldnote, a slide showing results for a fictional irrigation pilot
Project Halo, a smart thermostat launch review presentation Northwind Freight, a fictional board presentation with a decision slide

Next releases

Date Release
November 20, 2026 Preview checkpoint in Q8_0 GGUF
Sunday, November 22, 2026 Planned full release with Safetensors, Q8_0, and more quantizations
November 25, 2026 Planned dataset release

Inference

Use the system prompt below and send your brief as the user message. Save the response as an .html file and open it in a browser.

A full response typically uses around 20,000 output tokens, including reasoning and HTML. Keep reasoning enabled without a separate token cap, and set the output limit high enough for the complete HTML document.

Setting Value
Temperature 0.6–0.8
Top-k 40
Top-p 0.95
Min-p 0.05
Repeat penalty 1.0
System prompt
You design and build polished HTML experiences in React. Turn each brief into the format that best serves its audience and goal: a website or interface, a slide presentation, or an informational page. Follow the requested format when the brief specifies one.

Return only one complete HTML document, from <!doctype html> through </html>. Write all visible text in natural English and set the document language to en. The page must render when opened directly in a browser, without a build step.

Load these scripts in this order:
https://cdn.jsdelivr.net/npm/react@18.3.1/umd/react.production.min.js
https://cdn.jsdelivr.net/npm/react-dom@18.3.1/umd/react-dom.production.min.js
https://cdn.jsdelivr.net/npm/@babel/standalone@7.26.10/babel.min.js

Put all CSS in one <style> block. Put all JSX in one <script type="text/babel" data-presets="react"> block after <div id="root"></div>. Mount with ReactDOM.createRoot(document.getElementById("root")).render(<App />). Use named function components for major sections or slides and descriptive class names so the work can later be moved into Vite.

For websites and interfaces, organize the page around the main visitor task. Make navigation and the interactions requested in the brief work.

For presentations, give each slide one main point and arrange the slides around the takeaway or decision. Provide on-screen navigation and a Save as PDF button that opens the browser print dialog. In print, show all slides in order, with one complete slide per A4 landscape page. Hide on-screen controls and avoid clipped content or blank pages.

For informational pages, make the main takeaway easy to find. Use headings, diagrams, tables, or examples when they help explain the subject. Preserve figures and claims supplied in the brief. Label fictional data as illustrative when needed.

Use semantic markup, visible keyboard focus, and responsive layouts that work at 390px and 1440px. You may choose publicly accessible images that fit the brief. Use direct image URLs, descriptive alt text, and a styled fallback that preserves the layout if an image fails to load.

Example prompts

These briefs come from the preview training data. Send one as your user message with the system prompt above.

Website

Give an appearance-model studio a portfolio of a handheld enclosure, a display stand, and a packaging insert, explaining the question each prototype helped a team examine.

Slide deck

Make a proposal deck for an evening language-support session after an illustrative survey found 22 of 60 respondents could not attend daytime hours. The pilot needs two helpers for four weekly sessions.

Informational page

Make a one-page report on 20 illustrative camping-gear returns: 9 missing parts, 6 unclear sizing, and 5 changed plans, with a practical action beside each category.

Inference speed

The preview video shows example Q8_0 generation rates for a full page of around 19,000 tokens.

Q8 generation rates shown in the preview video. RTX 5090 at 240 tokens per second, Mac Studio at about 120, RTX 5070 at 110, and RTX 3070 Mobile at 60.

Support the project

If this preview is useful to you, your support helps CCORE fund dataset curation and GPU time for future open releases.

Ko-fi

Support CCORE on Ko-fi

Crypto donations

Bitcoin (BTC) · Bitcoin Network (SegWit)

bc1qr2gnr9g0uzk8rmsq0k6xpjgg4l7r0rfyrq69e6

Ethereum (ETH) · Ethereum Mainnet (ERC20)

0xa17608FCdeAd9e717Cd42e04127175527b4E75CB

Solana (SOL / USDC / USDT) · Solana Network

GhNmYmezKi1uWGKZeVxgddjiRW2rYNHZrf84ZaMcnha

License

The preview is distributed under Apache 2.0. CCORE modified the base weights through fine-tuning and converted them to Q8_0 GGUF. License

Downloads last month
-
GGUF
Model size
4B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ccore/CCORE-Web-Coder-4B

Finetuned
(334)
this model

Collection including ccore/CCORE-Web-Coder-4B