Nanbeige 4.2 Local Agent Studio

📺 Watch the full tutorial on YouTube

⚡ Overview

Nanbeige 4.2 Local Agent Studio integrates directly with Nanbeige/Nanbeige4.2-3B from Hugging Face. Utilizing its 3B non-embedding parameter Looped Transformer architecture, it runs agentic multi-step tool calls, structured reasoning traces (<think> tags), and automated report compilation.

🔄 Agent Execution Architecture

 ┌────────────────────────────────────────────────────────┐
 │                 👤 User Input Prompt                   │
 └───────────────────────────┬────────────────────────────┘
                             │
                             ▼
 ┌────────────────────────────────────────────────────────┐
 │       📝 Format ChatML Prompt with <think> Tag         │
 └───────────────────────────┬────────────────────────────┘
                             │
                             ▼
 ┌────────────────────────────────────────────────────────┐
 │    🧠 Nanbeige 4.2-3B Model Inference (Hugging Face)   │
 └───────────────────────────┬────────────────────────────┘
                             │
                             ▼
 ┌────────────────────────────────────────────────────────┐
 │       🔍 Regex Parser (parse_nanbeige_response)        │
 └───────┬───────────────────┼────────────────────┬───────┘
         │                   │                    │
         ▼                   ▼                    ▼
┌──────────────────┐ ┌────────────────┐ ┌──────────────────┐
│💭 Reasoning Trace│ │ 🛠️ XML Tool    │ │💬 Final Answer   │
│   (<think>)      │ │    Payload     │ │     Output       │
└────────┬─────────┘ └───────┬────────┘ └────────┬─────────┘
         │                   │                   │
         └───────────────────┼───────────────────┘
                             │
                             ▼
 ┌────────────────────────────────────────────────────────┐
 │      📊 Render Web Dashboard & Export outputs.md       │
 └────────────────────────────────────────────────────────┘

🚀 Quickstart & Installation

1. Download Model from Hugging Face

Download the official model directly from Hugging Face:

pip install huggingface_hub transformers torch accelerate
huggingface-cli download Nanbeige/Nanbeige4.2-3B

2. Environment Setup

python -m venv venv
.\venv\Scripts\Activate.ps1
pip install -r requirements.txt

3. Run Studio

python app.py

Open http://127.0.0.1:5000 in your web browser.

💻 Hugging Face Python Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Nanbeige/Nanbeige4.2-3B"

tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    use_fast=False,
    trust_remote_code=True
)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)

🛠️ Tech Stack

• Core Model: Nanbeige/Nanbeige4.2-3B (Hugging Face Transformers)
• Backend Framework: Python 3.12, Flask, Torch, Transformers, Jinja2
• Agent Scaffolding: XML Tool Parser, Multi-Turn Thinking Trace (enable_thinking, preserve_thinking)
• Frontend UI: Vanilla CSS3 (Glassmorphic dark design), Modern HTML5, Asynchronous JavaScript

📂 File Architecture

🎯 5 Key Use Cases

1. 🤖 Local Code Agent: Automated bug detection, refactoring, and test-case generation using SWE-Bench verified scaffolds.
2. 📄 Office PDF & Document QA: Zero-hint rubric analysis over dense document collections.
3. 🔍 Deep Research Assistant: Multi-step iterative search, synthesis, and structured report compilation.
4. 🛠️ XML Tool Call Pipeline: Seamless integration with local execution tools, shell commands, and API endpoints.
5. 🧠 Stream-of-Thought Audit: Audit internal reasoning steps (<think> tags) prior to tool execution for explainable AI safety.

🔮 5 Future Roadmap Features

1. 🌐 Native MCP (Model Context Protocol) Server Integration for live browser & OS automation.
2. 📊 Direct Benchmark Evaluator Interface to run GDPval and LiveCodeBench suites locally.
3. ⚡ vLLM & SGLang High-Throughput Inference Engine Toggle for batch processing.
4. 🔄 LoopSplit Visualizer step-by-step depth attention layer inspection.
5. 💾 Persistent Session Storage with vector retrieval over local knowledge bases.

🧪 Testing & Verification

Run the automated test runner:

python app.py --test

🔑 Keywords

Nanbeige 4.2 Nanbeige 3B Local AI Agentic LLM Hugging Face Looped Transformer XML Tool Calling SWE-Bench Offline AI PyTorch Flask AI Studio