Nanbeige4.3-33B

Nanbeige4.3-33B

A 33.1B open model for agentic coding and professional cowork tasks, built as the flagship teacher of the Nanbeige4.3 series. For developers and agent builders who want a strong open model near 30B parameters.

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: Nanbeige4.3-33B is a text generation model with 33.1B parameters. The authors built it as the flagship teacher of the Nanbeige4.3 series, to carry agent skills into smaller models later. It targets complex coding and professional cowork tasks, and it speaks English and Chinese.

Who it suits: It suits developers and agent builders who want an open model of about 30B parameters for coding agents and office-style task automation.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a coding agent on a repository | Fix a bug found by a failing test
Office automation | Operations teams | Automate multi-step professional tasks | A task that touches several tools in a row
AI research | Researchers | Study teacher models for distillation | Use it as the source for a smaller student model
Enterprise IT | Platform teams | Host an open model for internal agents | An internal assistant that plans and calls tools

If you need a small model for a phone or laptop, you can skip it, since the card plans a smaller 3B model for that.

Where to use it:

• Python with the transformers library
• Agent setups for coding and tool use, such as the ones the card used for its tests
• Windows, macOS or Linux machines with enough GPU memory for a 33B model

Use case | Who | What they get back
Agentic coding | Developers | Code changes and answers about a repository
Professional cowork tasks | Operations teams | Planned, multi-step work done through tools
Student model teaching | Researchers | A teacher whose outputs can train smaller models

How to set it up: You need these first:

• Python and pip
• A GPU setup with room for about 66.2 GB of BF16 weights, or about 16.6 GB at a 4-bit estimate

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Load the model with transformers as the model card describes.

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

How to use it: Chat with it through the model's chat template, or wire it into an agent. The card's tests ran it at its highest thinking effort with a chat template setting that keeps the thinking text.

Tool | What it does
preserve_thinking=true | A chat template setting that keeps the model's thinking text, used in the card's tests
Highest thinking effort | The mode in which the card compared all models

Example: You ask it to fix a failing test in a project through your coding agent. The model reads the code, plans the change, calls tools and returns a patch. Its Terminal-Bench 2.1 result in the card is 80.9.

For newcomers and experts: This is a large model for building agents. It is not a quick download, so a good first step is to read the model card and decide whether your GPU can hold the 4-bit version.

The card compares it with three rivals of similar size, Gemma4-31B, Qwen3.8-27B and Muse-Glimmer-30B, with the same inference settings for each. It leads on most agentic coding and cowork tables, but not on SWE-Bench Pro, GDPval or GPQA Diamond.

You give the agent a task
The model plans and thinks
It calls tools and reads results
It returns code or a finished answer

Good to know: These are the authors' own scores, and each benchmark has its own setup noted in the card.

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

• Where it trails: SWE-Bench Pro 60.3 against 61.7 for Qwen3.8-27B, GDPval 84.8 against 86.4, and GPQA Diamond 89.5 against 91.3 for Gemma4-31B
• Teacher role: the authors say it will feed a later distillation pipeline for smaller models such as Nanbeige4.3-3B
• Other versions: 101 quantized and 11 fine-tuned versions made by other people are listed
• License: Apache 2.0
• Cost to run: none named, you supply the GPU
• Data it reaches: the prompts and any files or tools you connect to it

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 33.1B
• Weights: about 66.2 GB in BF16; about 16.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 21,843
• Likes: 1,077
• License: apache-2.0
• Languages: en, zh
• Based on: Nanbeige/Nanbeige4.3-33B-Base
• Quantized versions: 102
• Fine-tunes: 12
• Tags: transformers, safetensors, llama, text-generation, nanbeige, conversational

Links

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

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