Nanbeige4.3-33B-GGUF-UltraOptimised-Curated-Protium

Nanbeige4.3-33B-GGUF-UltraOptimised-Curated-Protium

Offers GGUF quantizations of the 33.1B Nanbeige4.3 model for llama.cpp and Ollama, with a reasoning protocol and up to 524,288 tokens of context. For developers who run coding models locally.

Made by SolsticeAI 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 a set of GGUF files for Nanbeige4.3-33B, a 33.1B dense open-weight model from Nanbeige LLM Lab, prepared by Solstice-AI. The files were quantized with importance matrices, and the release adds a reasoning protocol the author calls Protium-1.0. The card aims it at software engineering, multi-file refactoring and terminal automation. It lists English and Chinese.

Who it suits: It suits developers who want a large coding model on their own hardware and have a strong GPU or plenty of system RAM.

Industry | Who uses it | How they use it | Example
Software engineering | Developers | Refactor code that spans many files with a local model | Loading a long repository into the large context window
DevOps | Automation builders | Drive command-line tasks from a local model | Letting a model plan and run shell steps
Local AI | Enthusiasts with a 24 GB GPU | Run the Q4_K_M file on a single graphics card | Serving it with llama-server
Research | Quantization researchers | Compare quant levels of one dense model | Testing Q4_K_M against Q8_0 on coding prompts

If your machine has less than 24 GB of GPU memory and little system RAM, you can skip it.

Where to use it:

• llama.cpp, through llama-server
• Ollama, through a Modelfile
• Windows, macOS or Linux, for downloading the files
• Apple Silicon Macs with 64 GB or more of unified memory, which the card suggests for the Q6_K file

Use case | Who | What they get back
Local coding assistant | Developers | Code and explanations from a model on your own machine
Whole-repository analysis | Engineers | Answers over a context window of up to 524,288 tokens
Inference server | Teams | A local server that other tools can call

How to set it up:

• Python and the Hugging Face hub package, for downloading
• llama.cpp or Ollama, to run the file
• Memory that fits your quant, as the table below shows

1. Install the Hugging Face hub package.
2. Download the model files.
3. Pick a quant that fits your hardware.
4. Start llama-server, or create an Ollama model from a Modelfile.

pip install -U huggingface_hub
hf download SolsticeAI/Nanbeige4.3-33B-GGUF-UltraOptimised-Curated-Protium

Quant | File size | Hardware the card suggests | Use
Q8_0 | 32.78 GB | 48 GB or more of VRAM, or 64 GB or more of system RAM | Reference tier with the highest precision
Q6_K | 25.31 GB | 32 to 48 GB of VRAM, or 64 GB of unified memory | Near-lossless tier for Apple Silicon and multi-GPU setups
Q5_K_M | 21.89 GB | 24 to 32 GB of VRAM, or 32 GB of RAM | Balanced choice with faster generation
Q4_K_M | 18.67 GB | A single 24 GB GPU such as an RTX 3090 or 4090 | The card's recommended default

How to use it: Start llama-server with the Q4_K_M file, or use the Ollama Modelfile. The card's sampling values are a temperature of 0.6 and a top-p of 0.95.

# Recommended baseline: Q4_K_M on 24GB GPU
llama-server \
  --model nanbeige4.3-33b-q4_k_m.gguf \
  --host 0.0.0.0 \
  --port 8080 \
  --ctx-size 65536 \
  -ngl 99 \
  --temp 0.6 \
  --top-p 0.95

For Ollama, create a Modelfile with these lines. The card's Modelfile may hold more lines than the ones shown here.

FROM ./nanbeige4.3-33b-q4_k_m.gguf

PARAMETER temperature 0.6
PARAMETER top_p 0.95
PARAMETER num_ctx 65536
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"

Tool | What it does
llama-server | Serves the GGUF file over HTTP on the port you set
--ctx-size | Sets the context window, 65536 tokens in the card's example
-ngl | Sets how many layers go to the GPU, 99 in the example
Modelfile | Tells Ollama which file to load and which settings to use

Example: You start llama-server with the Q4_K_M file on a 24 GB GPU, then send it a request to refactor a function used in several files. Text comes back, and the card says reasoning may appear inside think blocks first.

For newcomers and experts: GGUF is a file format that llama.cpp and Ollama read. A quant stores the model with fewer bits, so the file shrinks and quality can drop. A good first try is the Q4_K_M file with the llama-server command above.

Under the hood, the model is a dense 33.1B network with 52 layers and a native context of 524,288 tokens. The example commands use 65536 to save memory. The card warns that quantizing below 4 bits harms multi-turn tool calling and syntax generation, so it treats Q4_K_M as the minimum for serious coding.

Nanbeige4.3-33B original model
Importance-matrix quantization to GGUF
Protium-1.0 reasoning protocol
Run in llama-server or Ollama

Good to know:

Title: File size of each quant in GB
Q8_0: 32.78 GB
Q6_K: 25.31 GB
Q5_K_M: 21.89 GB
Q4_K_M: 18.67 GB

• A Q3_K_M file of 15.2 GB sits in a separate repository, and the card advises against it for agentic software engineering or complex math.
• The card's claims of near-lossless quality come without published benchmark numbers.
• The Protium-1.0 notes describe think blocks to stop circular reasoning, and an answer-first habit before the step-by-step checks.
• License: Apache 2.0
• Cost to run: none named, though the card expects a 24 GB GPU or more for the recommended file
• Data it reaches: none named, since it runs from local files in llama.cpp or Ollama
• Not for: the Q3_K_M file in agentic software engineering or complex math, and machines with too little memory for these sizes

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 33.1B
• Weights: about 16.6 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 767
• Likes: 6
• License: apache-2.0
• Languages: en, zh
• Based on: Nanbeige/Nanbeige4.3-33B
• Tags: gguf, solstice-ai, nanbeige, nanbeige4.3, nanbeige4.3-33b, boss-zhipin, kanzhun, protium, protium-1.0, cognitive-architecture, dynamic-reasoning, objective-gated

Links

• Model: https://huggingface.co/SolsticeAI/Nanbeige4.3-33B-GGUF-UltraOptimised-Curated-Protium
• Files: https://huggingface.co/SolsticeAI/Nanbeige4.3-33B-GGUF-UltraOptimised-Curated-Protium/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download SolsticeAI/Nanbeige4.3-33B-GGUF-UltraOptimised-Curated-Protium

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