UIGEN-X-30B-MoE

Unsloth-tuned Qwen3 30B mixture-of-experts model for UI generation, heavy coding, reasoning, and agentic workflows.

🚀 Overview

UIGEN-X-30B-MoE is a GGUF distribution of Tesslate/UIGENT-30B-3A-Preview, an Unsloth-tuned, UI-generation-focused mixture-of-experts model built on Qwen/Qwen3-Coder-30B-A3B-Instruct.

The model uses a sparse MoE design: of its 30.5B total parameters, only 3.3B are activated per token (top 8 of 128 routed experts across 48 layers). This delivers the quality of a large model with the inference cost of a small one, making it practical to run locally for complete front-end UI generation, multi-step coding, and tool-using agent workflows. It natively supports a 256K-token context (extendable toward 1M with YaRN).

🎯 Key Features

• UI generation: Systematically plans, architects, and implements complete user interfaces across modern web stacks (HTML, Tailwind CSS, React, and more)
• Sparse MoE efficiency: 30.5B total parameters but only 3.3B active per token (8 of 128 experts), keeping inference fast and memory-friendly
• Long context: 256K native context window for large repositories, documents, and multi-file projects
• Coding & reasoning: Strong multi-step code generation and structured reasoning inherited from the Qwen3-Coder base
• Agentic / tool use: Tuned for function-calling and tool-invocation patterns in agent systems
• Full local tooling: Ships with tokenizer metadata and chat template for drop-in use with Ollama

🏷️ Available Versions

Tag casing matches the published tags exactly. Pull commands are case-sensitive.

💻 Quick Start

# Recommended quantization (Q4_K_M)
ollama run richardyoung/uigen-x-30b-moe:Q4_K_M

# Smallest footprint
ollama run richardyoung/uigen-x-30b-moe:q2_k

# Best quality
ollama run richardyoung/uigen-x-30b-moe:q8_0

# Example: generate a UI
ollama run richardyoung/uigen-x-30b-moe:Q4_K_M "Build a responsive pricing page with three tiers using Tailwind CSS."

🛠️ Use Cases

• UI / front-end generation: Produce complete, styled interfaces in HTML, Tailwind, React, and other modern stacks
• Code generation: Multi-step coding across files and frameworks
• Agentic workflows: Tool-calling and function invocation for autonomous agents
• Reasoning tasks: Structured, step-by-step problem solving
• Large-context analysis: Review long documents, PDFs, and full repositories

📋 System Requirements

Because of the sparse MoE design, only ~3.3B parameters are active per token, so effective throughput is much faster than a dense 30B model at the same size.

🔧 Technical Details

Source Model: Tesslate/UIGENT-30B-3A-Preview Base Model: Qwen/Qwen3-Coder-30B-A3B-Instruct Architecture: Qwen3 MoE (Qwen3MoeForCausalLM) Parameters: 30.5B total / 3.3B active per token Experts: 128 routed experts, 8 activated per token Layers: 48 Context Length: 262,144 tokens (256K), extendable toward 1M with YaRN Quantization: GGUF (q2_k through q8_0) via llama.cpp License: Apache 2.0

⚠️ Disclaimer

This model is provided as-is under the Apache 2.0 license. Generated UI code, application logic, and other output may contain errors, security issues, or accessibility gaps and must be reviewed and tested before any production use. The author of this distribution is not affiliated with Tesslate or the Qwen team and makes no warranty as to fitness for any purpose. Use responsibly and in accordance with applicable laws and the upstream model licenses.

🙏 Acknowledgments

• Source Model: Tesslate/UIGENT-30B-3A-Preview by Tesslate
• Base Model: Qwen/Qwen3-Coder-30B-A3B-Instruct by Alibaba (Qwen)
• Fine-tuning: Unsloth
• Quantization: llama.cpp

Built & maintained by Richard Young · DeepNeuro