ViCoder-html-32B-preview

🚀 A powerful HTML/CSS/JS sketching model powered by Qwen2.5-Coder-32B-Instruct 🚀

Developed by Vichar AI | Hugging Face Profile
Licensed under Apache 2.0

💡 What is ViCoder-html-32B-preview?

ViCoder-html-32B-preview is a preview model in the ViCoder series from Vichar AI — a line of models specialized in code generation. This model focuses specifically on sketching single-page websites, such as landing pages and dashboards, using using:

🧠 HTML for semantic structure
🎨 Tailwind CSS for modern, utility-first styling
⚙️ JavaScript for interactivity and basic dynamic behavior

This model is ideal for:

Web Developers: Quickly scaffolding dashboards or page layouts.
Frontend Engineers: Prototyping UIs and exploring design variations.
Designers: Turning textual mockups into initial code sketches.
Educators & Students: Learning and experimenting with HTML, Tailwind CSS, and JavaScript in a practical context.

⚠️ Note: This is a preview version. It demonstrates core capabilities but is still under active development. A more refined and robust production release is planned. Stay updated via vichar.io or follow VicharAI on Hugging Face!

🛠️ Model Details

Property	Value
Model Type	Code Generation (Instruction-tuned Language Model)
Base Model	Qwen/Qwen2.5-Coder-32B-Instruct
Developed by	Vichar AI (HF Profile)
Languages	Primarily HTML, Tailwind CSS, JavaScript. Understands English instructions.
Training Data	Proprietary curated dataset focusing on high-quality web components and pages.
License	Apache 2.0
Library	🤗 Transformers
Contact	Visit vichar.io or use HF Discussions

🧱 GGUF Quantized Versions

Quantized versions of ViCoder-html-32B-preview in GGUF format are available for efficient local inference using llama.cpp, LM Studio, or Ollama.

You can find them here:

🔗 GGUF Quantizations on Hugging Face

These quantized variants (Q3_K_M, Q4_K_M, Q6_K, Q8_0) are useful for running the model on lower-memory hardware or for embedding in desktop/web applications.

⚡ Example Usage

Use the transformers library pipeline for easy text generation. Ensure you have transformers, torch, and accelerate installed.

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import torch

# Define model ID
model_id = "VicharAI/ViCoder-html-32B-preview"

# Load tokenizer and model
# Use bfloat16 for faster inference if your GPU supports it
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16, # Or torch.float16 if bfloat16 is not supported
    device_map="auto"           # Automatically distribute across available GPUs/CPU
)

messages = [
    {"role": "user", "content": "A modern, sleek landing page for a company focusing on open-source LLM solutions"},
]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize = True,
    add_generation_prompt = True,
    return_tensors = "pt",
).to("cuda")

from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer, skip_prompt = True)
_ = model.generate(input_ids = inputs, streamer = text_streamer, max_new_tokens = 16000,
                   use_cache = True, temperature = 0.7, min_p = 0.1, repetition_penalty=1.1)

✨ Output Sample

<!DOCTYPE html>
<html lang="en">
<head>
  <meta charset="UTF-8">
  <title>Our Love Story - Surprise Website</title>
  <meta name="viewport" content="width=device-width, initial-scale=1.0">
  <!-- Tailwind CSS CDN -->
  <script src="https://cdn.tailwindcss.com"></script>
  <style>
    /* Custom animation classes */...

(Note: The model aims to generate complete HTML structures with Tailwind classes. Review and adapt generated code as needed.)

🧪 Evaluation & Limitations

As a preview release, this model has undergone initial internal testing focused on:

Code Correctness: Validity of generated HTML, Tailwind CSS classes, and basic JavaScript snippets.
Tailwind CSS Usage: Adherence to Tailwind's utility-first principles and common patterns.
Component Structure: Logical organization of HTML elements for typical web components.
Instruction Following: Ability to understand and implement requirements from the prompt.

Current Limitations:

No Formal Benchmarks: Not yet evaluated on standard code generation benchmarks (e.g., HumanEval-X, MBPP).
Complex Logic: May struggle with complex JavaScript logic, state management, or intricate CSS beyond Tailwind utilities.
Hallucination Risk: Like all LLMs, it can sometimes generate incorrect, incomplete, or non-optimal code. Always review the output.
Preview Status: Not recommended for critical production use without thorough validation.

📍 Roadmap

The ViCoder series is an ongoing project at Vichar AI. Our current roadmap includes:

✅ ViCoder-html-32B-preview: Initial public preview release (this model).
⏳ ViCoder-html-32B (v1.0): Planned production-ready release with improved training data, fine-tuning, and evaluation.
🚀 ViCoder-js-32B: Future model focusing specifically on advanced JavaScript generation (frameworks, logic).
🐍 ViCoder-python-32B: Potential companion model for Python backend code generation.
📊 Benchmarking & Evaluation: Formal evaluation on relevant code generation benchmarks.

Follow VicharAI on Hugging Face or check the Vichar AI website for announcements!

📄 License

This model and its code are licensed under the Apache License 2.0. You can find the full license text here.

🙏 Citation

If you use ViCoder-html-32B-preview in your projects, publications, or research, please cite it:

@misc{vicharai_vicoder_html_32b_preview_2025,
  title        = {ViCoder-html-32B-preview: A Preview Model for HTML/Tailwind CSS/JavaScript Sketching},
  author       = {Vichar AI},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/VicharAI/ViCoder-html-32B-preview},
  url          = {https://vichar.io}
}

📬 Get in Touch

We welcome feedback, questions, and collaboration ideas!

Hugging Face: Open an issue or start a discussion on the model page's Community tab.
Website: Visit us at https://vichar.io for more information about Vichar AI and the ViCoder project.
Contact: Find direct contact methods on the Vichar AI website.

🤝 Acknowledgments

This project builds upon the incredible work of others:

SprykAI for their support during model experimentation phases.
The Qwen Team at Alibaba Cloud for developing the foundational Qwen2.5-Coder-32B-Instruct model.
The Hugging Face Team for their platform and libraries (🤗 Transformers, Accelerate,TRL).
The broader open-source AI community for continuous innovation and shared knowledge.
Development efforts by the team at Vichar AI.

💥 This preview is just the start! Explore, build, and stay tuned for the full ViCoder suite from Vichar AI! 💥

vicharai
/

ViCoder-html-32B-preview