fnruha0921/VibeThinker-3B-Fable5-UI

VibeThinker-3B: Agentic Frontend Specialist

VibeThinker-3B is a fine-tuned version of Qwen2.5-3B, specifically engineered for Agentic Frontend Development. Unlike the base model, VibeThinker-3B is trained to handle complex UI/UX tasks, follow agentic reasoning traces (Claude-style), and perform Tool Calling for automated coding workflows.

Model Details

Model Description

VibeThinker-3B is designed to bridge the gap between simple code generation and autonomous agentic behavior. It excels at creating production-ready frontend components using modern stacks (React, Next.js, Tailwind CSS, shadcn/ui) while maintaining a high level of reasoning for complex logic.

• Developed by: WeiboAI (Finetuned for Fable Frontend Project)
• Model type: Causal Language Model (LoRA Finetune)
• Language(s): English, Korean
• Finetuned from model: Qwen/Qwen2.5-3B
• Key Focus: Tool Calling, Agentic Reasoning, Frontend Orchestration

Model Sources

• Repository: HuggingFace: WeiboAI/VibeThinker-3B

Uses

Direct Use

• Agentic Coding: Generating step-by-step reasoning for complex features before implementation.
• Frontend Generation: Creating high-quality React/Next.js components with best practices.
• Tool Calling: Interacting with external APIs or local functions to automate development tasks.
• UI/UX Prototyping: Rapidly building functional UI from natural language descriptions.

Out-of-Scope Use

• Competitive programming (LeetCode style) is supported but not the primary focus.
• Heavy backend-only logic without frontend context.

Bias, Risks, and Limitations

• Tool Calling Precision: While trained for tool use, users should verify the generated JSON arguments for safety.
• Context Window: Optimized for 8,192 sequence length. Performance may degrade for extremely large codebases without RAG.

How to Get Started with the Model

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "WeiboAI/VibeThinker-3B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

messages = [
    {"role": "user", "content": "Create a responsive dashboard using Next.js and shadcn/ui."}
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")
outputs = model.generate(inputs, max_new_tokens=2048)
print(tokenizer.decode(outputs[0]))

Training Details

Training Data

The model was trained on a curated mix of ~11,000 high-quality samples:

1. Stage 1: Agentic Reasoning (40%): Claude-style reasoning traces, CodeAlpaca, and Magicoder-Evol datasets.
2. Stage 2: Frontend Specialization (40%): Specialized Next.js, React, and shadcn/ui component datasets.
3. Stage 3: Tool Calling & Agentic (20%): Glaive Function Calling v2 and Hermes Function Calling datasets.

Training Procedure

• Method: LoRA (Low-Rank Adaptation)
• Sequence Length: 8,192
• Rank (R): 64
• Alpha: 128
• Learning Rate: 1e-4
• Epochs: 3
• Optimizer: AdamW with cosine schedule

Technical Specifications

• Hardware: Trained on Together AI Cloud Infrastructure.
• Framework: PEFT 0.15.1, Transformers.

Citation

BibTeX:

@misc{vibethinker2026,
  author = {WeiboAI Team},
  title = {VibeThinker-3B: Agentic Frontend Specialist},
  year = {2026},
  publisher = {HuggingFace},
  journal = {HuggingFace Model Hub}
}