Nanbeige4.3-33B-INT4

Nanbeige4.3-33B-INT4

INT4 quantized version of Nanbeige4.3-33B, a 33B model for coding and professional agent tasks, made with quantization-aware training. For developers who want a strong open agent model in less memory.

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: This is the INT4 version of Nanbeige4.3-33B, a 33 billion parameter model aimed at coding and professional office-style agent tasks. The makers used quantization-aware training, including distillation from the BF16 model and reinforcement learning on the INT4 model itself, to keep quality at 4-bit. It works in English and Chinese.

Who it suits: It suits developers who want a strong open agent model in the 30B class that runs in less memory than the BF16 original.

Industry | Who uses it | How they use it | Example
Software | Agent developers | Run a coding and tool-using agent locally | Fixing issues in a repository with tools
Office work | Automation builders | Handle professional cowork tasks | Running an automation benchmark task
Research | ML researchers | Study quantization-aware training | Comparing INT4 behavior with the BF16 model

You can skip it if you want a tiny model, since 33 billion parameters still need a serious GPU.

Where to use it: It runs in Python through the transformers library on Windows, macOS or Linux. The card gives no run commands in the visible part.

• Python scripts that use transformers
• Agent frameworks that call a local model
• A GPU machine with enough memory for a 33B model in 4-bit

Use case | Who | What they get back
Agentic coding | Developers | A model tested on terminal and repository tasks
Professional cowork | Automation teams | A model tested on job, office and planning tasks
Reasoning | Researchers | A model tested on science and math questions

How to set it up: You need Python and the Hugging Face tool. Hugging Face's own record lists about 132 GB counted in 32-bit integers, which does not show the real 4-bit file size, so check the file sizes before you download. The card's visible part gives no run steps.

• Python and the Hugging Face tool
• A GPU with enough memory for a 33B model in 4-bit
• Disk space, after checking the real file sizes

1. Install the Hugging Face download tool.
2. Download the repository files.
3. Load the model with transformers and follow the model card for details.

pip install -U huggingface_hub
hf download Nanbeige/Nanbeige4.3-33B-INT4

How to use it: Load the model through transformers and use it as a chat or agent model. The BF16 results on the card were measured with preserve_thinking set to true in the chat template, and the models ran at their highest thinking effort.

Tool | What it does
hf download | Fetches the model files to your computer
transformers | Loads the model for use in Python

Example: You connect the model to a terminal agent and ask it to fix a failing test in a repository. It plans, runs commands, edits files and reports the result, as the coding benchmarks on the card measure.

For newcomers and experts: If you are new, this is a large open model squeezed to 4 bits so it fits in less memory. Read the card first and start with simple chat before building an agent.

If you know the field, the 33B model is a teacher meant to feed distillation into smaller Nanbeige4.3 models. For INT4, on-policy distillation from the BF16 model and reinforcement learning on the INT4 model recover quality. The benchmark table on the card is for the BF16 model.

Nanbeige4.3-33B BF16 model
Quantization-aware training
Distillation from BF16 to INT4
Reinforcement learning on INT4
INT4 model for local use

Good to know:

• License: Apache 2.0
• Cost to run: no API key, account or paid service is named, but you need a GPU for a 33B model
• Data it reaches: the prompts you give it on your machine, plus files downloaded from Hugging Face
• Not for: treating the BF16 scores as proof for this INT4 copy

Title: Terminal-Bench 2.1 score for the BF16 model
Nanbeige4.3-33B: 80.9
Qwen3.8-27B: 73.0
Muse-Glimmer-30B: 61.8
Gemma4-31B: 47.2

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 33.1B
• Weights: about 132 GB in I32
• Formats: Safetensors
• Downloads in the last 30 days: 749
• Likes: 10
• License: apache-2.0
• Languages: en, zh
• Based on: Nanbeige/Nanbeige4.3-33B
• Tags: transformers, safetensors, llama, text-generation, nanbeige, conversational, compressed-tensors

Links

• Model: https://huggingface.co/Nanbeige/Nanbeige4.3-33B-INT4
• Files: https://huggingface.co/Nanbeige/Nanbeige4.3-33B-INT4/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Nanbeige/Nanbeige4.3-33B-INT4

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