Description
Nanbeige4.2-3B is a compact model developed to deliver strong agentic behavior and broad reasoning capabilities. Built on the Nanbeige4.2-3B-Base, it utilizes a Looped Transformer architecture that reuses transformer layers to enhance model capacity without increasing parameters. With only 3 billion non-embedding parameters, it achieves impressive performance in general-agent and code-agent tasks.
Read More
The model undergoes supervised fine-tuning (SFT) to diversify training environments and task types, ensuring high-quality training data through rigorous filtering and validation processes. During reinforcement learning (RL), outcome and process rewards are combined to improve training stability.
Nanbeige4.2-3B demonstrates solid agentic behavior across complex tool-use, office-agent, and code-agent benchmarks, outperforming larger models like Qwen3.5-9B and Gemma4-12B. It also excels in mathematical, coding, and scientific reasoning tasks, continuing the strong reasoning performance of its predecessor, Nanbeige4.1-3B.
The model is suitable for local personal assistant applications, supporting extended tasks in daily assistance, office work, and deep research when integrated with an agentic scaffold like OpenClaw. It is compact enough for local deployment while retaining the necessary capabilities for multi-step workflows.
Nanbeige4.2-3B supports a context length of up to 262,144 tokens and offers configurable chat templates for reasoning and tool-use scenarios. Users can adjust inference settings based on the target scenario, such as agentic tasks or reasoning and chat tasks.
Despite its strengths, the model may generate unexpected or inappropriate outputs due to its probabilistic nature. Users are advised to handle such outputs responsibly, as the developers do not assume responsibility for any consequences arising from inappropriate content.
- CategoryAI Models & LLMs
- Pricing ModelOpen Source
- DeploymentRuns locally
- PlatformsWeb
Nanbeige4.2-3B Highlights
Looped Transformer architecture
3 billion non-embedding parameters
Strong agentic behavior
Broad reasoning capabilities
Supervised fine-tuning
Reinforcement learning stability
Local personal assistant suitability
Configurable chat templates
Getting Started with Nanbeige4.2-3B
Access model page: Visit Hugging Face
Authenticate: Log in to your account
Configure: Set up chat templates
Send prompts: Interact with the model
Fine-tune: Adjust settings for tasks
Nanbeige4.2-3B's Use Cases
- General-Agent Tasks
- Code-Agent Tasks
- Mathematical Reasoning
- Scientific Research
- Personal Assistance
