Nanbeige4.2-3B by Nanbeige

Nanbeige4.2-3B by Nanbeige

Compact looped-transformer model of 4.2B parameters for agent work such as tool use, office tasks and coding, in English and Chinese. For developers who want a small open model for agents.

Made by Nanbeige on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Nanbeige4.2-3B is a compact language model for agent work, built on Nanbeige4.2-3B-Base. It uses a looped transformer that reuses its layers to add capacity without adding parameters. It has 4.2 billion parameters in total and about 3 billion outside the embeddings. It works in English and Chinese.

Who it suits: It suits developers who want a small open model for tool use, office tasks and coding agents on modest hardware.

Industry | Who uses it | How they use it | Example
Software | Agent developers | Power a tool-using or coding agent | Running a code agent on a bug-fix task
Office work | Productivity tool makers | Automate document and office tasks | Filling in an office workflow with tools
Research | ML researchers | Study looped transformer designs | Comparing it with larger open models
Personal tools | Hobbyists | Run a local assistant | Pairing it with an agent scaffold for daily tasks

You can skip it if you need a huge model for hard open-ended work, since this one is aimed at the 3B scale.

Where to use it: It runs in Python through the transformers library on Windows, macOS or Linux. The card suggests pairing it with an agent scaffold for personal workflows, naming OpenClaw as one.

• Python scripts that use transformers
• An agent scaffold such as OpenClaw, for personal assistant use
• A local computer, since the weights are small

Use case | Who | What they get back
Tool use | Agent builders | Tool-calling behavior on agent benchmarks
Code agents | Developers | Fixes tested on software engineering tasks
Local assistant | Hobbyists | Help with daily, office and research tasks

How to set it up: You need Python and the Hugging Face tool. Hugging Face lists the weights at about 8.3 GB in BF16, or an estimated 2.1 GB at 4-bit. The card's visible part gives no run commands, so follow its usage section.

• Python with transformers
• About 8.3 GB of disk for the BF16 weights
• A GPU for comfortable speed

1. Install the Hugging Face download tool.
2. Download the model files, or let transformers fetch them by name.
3. Load the model and tokenizer, and connect an agent scaffold if you want one.

pip install -U huggingface_hub
hf download Nanbeige/Nanbeige4.2-3B

How to use it: Load it with transformers for chat, or connect it to an agent scaffold so it can use tools. The repository also ships a Python modeling file that carries the makers' newest architecture changes. Three of them are named LoopSplit, depth-attention mHC and n-gram embeddings.

Tool | What it does
hf download | Fetches the model files to your computer
transformers | Loads the model for use in Python
Agent scaffold | Gives the model tools and a workflow, such as OpenClaw

Example: You connect the model to a code-agent scaffold and give it a bug report. It reads the code, calls tools, and proposes a fix, which the scaffold can test.

For newcomers and experts: If you are new, load the model in transformers and chat with it first. Add an agent scaffold once you are comfortable with the basics.

If you know the field, supervised tuning drew on real and synthesized environments, with filtering by test cases and rubrics. Reinforcement learning mixed outcome and process rewards to keep a small model stable. The looped design reuses layers to grow capacity.

Nanbeige4.2-3B-Base
Supervised tuning on agent environments
Reinforcement learning with outcome and process rewards
Compact agentic model

Good to know: The makers report higher agent benchmark scores than Gemma4-12B and Qwen3.5-9B, both bigger models. The figures are from their own comparison.

• License: Apache 2.0
• Cost to run: no API key, account or paid service is named
• Data it reaches: the prompts you give it on your machine, plus files downloaded from Hugging Face and whatever your agent scaffold lets it use
• Not for: results beyond what the makers' own benchmarks show

Title: SWE-Bench Verified score
Nanbeige4.2-3B: 63.6
Qwen3.5-9B: 53.1
Gemma4-12B: 44.2
Qwen3.5-4B: 38.8

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 4.2B
• Weights: about 8.3 GB in BF16; about 2.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 96,886
• Likes: 860
• License: apache-2.0
• Languages: en, zh
• Based on: Nanbeige/Nanbeige4.2-3B-Base
• Quantized versions: 60
• Fine-tunes: 18
• Tags: transformers, safetensors, nanbeige, text-generation, conversational

Links

• Model: https://huggingface.co/Nanbeige/Nanbeige4.2-3B
• Files: https://huggingface.co/Nanbeige/Nanbeige4.2-3B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Nanbeige/Nanbeige4.2-3B

Then follow the usage in its model card. Nothing in this kit runs by itself.