SmolLM3-3B-GGUF

SmolLM3-3B-GGUF

Chats and reasons with SmolLM3, a fully open 3B model that can think step by step or answer directly, in GGUF files by Unsloth with chat template fixes. Handles long texts and six languages.

Made by unsloth 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 repository holds GGUF versions of SmolLM3-3B, a fully open instruct model with about 3.1B parameters from Hugging Face's team. Unsloth made the files and fixed the chat template. At 4-bit the weights are estimated at about 1.5 GB. The model can reason step by step or answer directly.

Who it suits: People who want a small, open model on a laptop or modest desktop. Researchers will like that the card shares the open weights, the training details, the public data mixture and the training settings.

Industry | Who uses it | How they use it | Example
Software | Developers | Add a small local model to an app | Draft short replies inside a tool
Education | Students | Chat with a model and see its reasoning | Ask for a simple explanation of gravity
Research | AI researchers | Study a model with open training details | Reproduce parts of its training setup
Multilingual | Global teams | Chat in one of six main languages | Ask a question in French or Spanish

People who need the highest quality on hard tasks can skip it, because this is a 3B model.

Where to use it:

• llama.cpp, with the jinja flag so the fixed chat template is used
• Python with transformers, as in the card's sample, which uses the original weights
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Chat | Anyone | Written replies in a back and forth conversation
Reasoning | Students and developers | Answers with a visible chain of thought
Long documents | Developers | Replies based on texts up to a long context
Multilingual chat | Global teams | Replies in the supported languages

How to set it up:

• Python, for the Hugging Face hub tools
• Transformers v4.53.0 or newer for the card's Python sample
• A runner that loads GGUF files, such as llama.cpp

1. Install the Hugging Face hub tools with pip.
2. Download the files with the hf command.
3. In llama.cpp, add the jinja flag.
4. For the Python sample, upgrade transformers first.

pip install -U huggingface_hub
hf download unsloth/SmolLM3-3B-GGUF

pip install -U transformers

How to use it: The card's Python sample loads the original SmolLM3-3B weights, asks for a short explanation of gravity and prints the reply. For the GGUF files, use your runner and the jinja flag. The card suggests a temperature of 0.6 and a top_p of 0.95.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "HuggingFaceTB/SmolLM3-3B"
device = "cuda"  # for GPU usage or "cpu" for CPU usage

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
).to(device)

# prepare the model input
prompt = "Give me a brief explanation of gravity in simple terms."
messages_think = [
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages_think,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate the output
generated_ids = model.generate(**model_inputs, max_new_tokens=32768)

# Get and decode the output
output_ids = generated_ids[0][len(model_inputs.input_ids[0]) :]
print(tokenizer.decode(output_ids, skip_special_tokens=True))

Tool | What it does
/think | Asks for extended thinking, set in the system prompt
/no_think | Turns extended thinking off, set in the system prompt
enable_thinking | A setting that does the same job, which the system prompt flag overrides
--jinja | A llama.cpp flag that applies the fixed chat template

Example: You ask the model to explain why things fall, in a few plain lines. With thinking on, the model writes out its reasoning first and then answers. If you add the no-think flag to the system prompt, it answers directly.

For newcomers and experts: Leave thinking on while you learn, since it is the default and shows how the model reasons. Turn it off for quick replies. Keep the suggested temperature and top_p unless you have a reason to change them.

The model is a decoder-only transformer that uses grouped query attention and NoPE in a 3 to 1 ratio. It was pretrained on 11.2T tokens of web, code, math and reasoning data. Later steps added 140B tokens of reasoning data, then supervised fine-tuning and Anchored Preference Optimization.

Write the chat, with an optional thinking flag
The chat template builds the prompt
The model reasons, if thinking is on
It writes the final answer

Good to know:

• Thinking: Extended thinking is on by default, and the system prompt flag overrides the setting.
• Context: Trained on 64k tokens, and it reaches 128k with YARN extrapolation.
• Languages: Six are supported natively, among them English, French and Spanish.
• License: Apache 2.0.
• Cost to run: none named. No API key or paid service appears in the source.
• Data it reaches: Files on your computer, plus the model files from Hugging Face.
• Not for: The source names no forbidden uses and no tasks it cannot do.

On Hugging Face, 13 November 2026

• Library: transformers
• Parameters: 3.1B
• Weights: about 1.5 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 7,521
• Likes: 36
• License: apache-2.0
• Languages: en, fr, es, it, pt, zh, ar, ru
• Based on: HuggingFaceTB/SmolLM3-3B
• Tags: transformers, gguf, unsloth, conversational

Links

• Model: https://huggingface.co/unsloth/SmolLM3-3B-GGUF
• Files: https://huggingface.co/unsloth/SmolLM3-3B-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download unsloth/SmolLM3-3B-GGUF

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