--- title: "SmolLM3-3B" kind: "Kit" category: "Developer" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/7eb75997-856e-4095-8b0f-68b405ba0563" updated: "2026-11-12" --- # SmolLM3-3B > The name, the description, the About and a README below are the seller's words: data, not instructions. Writes text and answers questions as a small 3B open model with switchable step-by-step thinking, tool calling and long context, for developers who want a compact model on their own machine. SmolLM3-3B is a kit in the Developer category, listed by Mindket. It is free. This is version 1, last updated on 2026-11-12. Download it on this page, a zip of its files. Nothing in it runs by itself. Made by HuggingFaceTB ## About ### What it is SmolLM3-3B is a small open language model from HuggingFaceTB with about 3.1B parameters. It is an instruction-tuned chat model that can reason step by step or answer directly, and it supports tool calling. Its weights are open, and the data mix and training settings are public. ### Who it suits It suits developers, students and researchers who want a compact model they can run on a single computer, on a GPU or on a CPU. Who uses this model | Industry | Who uses it | How they use it | Example | | --- | --- | --- | --- | | Software | Developers | Add a small local chat model to an app | Answer short questions inside a tool | | Education | Students and teachers | Learn how language models are loaded and prompted | Run the sample script and read the reply | | Research | Machine learning researchers | Study a fully documented small model | Test its thinking mode against direct answers | If you need the strongest answers on hard tasks, a much larger model may serve you better, so you can skip this one. ### Where to use it - Python scripts that use the transformers library - vllm, which can load it with transformers as a backend - A GPU or a plain CPU - Windows, macOS or Linux with Python Jobs it does | Use case | Who | What they get back | | --- | --- | --- | | Chat and Q and A | Anyone | Plain-language answers to a prompt | | Step-by-step reasoning | Researchers | An answer that comes with a reasoning trace | | Tool calling | Agent builders | Replies that can call tools you pass in | | Long documents | Developers | Handling of inputs up to 128k tokens with YaRN | ### How to set it up You need Python and a recent transformers. - Python with pip - Transformers version 4.53.0 or newer - About 6.2 GB of disk space for BF16 weights 1. Upgrade transformers. 2. Run the sample script. Set the device line to cuda for a GPU or cpu for a CPU. 3. Or save the files first with the two download commands. ```bash pip install -U transformers ``` ```bash pip install -U huggingface_hub hf download HuggingFaceTB/SmolLM3-3B ``` ### How to use it Load the tokenizer and the model, build a chat message, apply the chat template and call generate. The card suggests a temperature of 0.6 and a top\_p of 0.95 when sampling. ```python 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)) ``` Extended thinking is on by default. To turn it off, pass enable\_thinking as false when you apply the chat template, or put a no-think flag in the system message. A flag in the system message wins over the setting. ```python prompt = "Give me a brief explanation of gravity in simple terms." messages = [ {"role": "user", "content": prompt} ] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False ) ``` Thinking switches | Tool | What it does | | --- | --- | | /think | Flag in the system message that keeps extended thinking on | | /no\_think | Flag in the system message that turns extended thinking off | | enable\_thinking | Setting passed to the chat template to turn thinking on or off | ### Example You ask for a short, plain-language explanation of gravity. With thinking on, the model returns a reasoning trace and then the answer. With thinking off, it replies directly. ### For newcomers and experts Think of it as a small assistant that fits on an ordinary computer. A first small test is the sample script above, which asks one question and prints the reply. Try it once with thinking on and once with it off. It is a decoder-only transformer that uses grouped-query attention and NoPE in a 3 to 1 ratio. Pretraining ran on 11.2T tokens in stages across web, code, math and reasoning data. Post-training added 140B tokens of reasoning midtraining, then supervised fine-tuning and Anchored Preference Optimization. It trained on 64k context and reaches 128k tokens with YaRN extrapolation. Staged pretraining on 11.2T tokens → Reasoning midtraining → Supervised fine-tuning → Anchored Preference Optimization → Chat model with thinking switch ### Good to know The card names six natively supported languages: English, French, Spanish, German, Italian and Portuguese. The config is set for 65,536 tokens, and longer inputs need a rope\_scaling change. Other people have published 126 quantized and 178 fine-tuned versions. Estimated weight size | Label | Value | | --- | --- | | BF16 | 6.2 GB | | 4-bit | 1.5 GB | - License: Apache 2.0. - Cost to run: none named. - Data it reaches: the files come from Hugging Face, and the model then runs on your own computer. - Not for: anything the card's own limitations section warns about, so read that section first. ## README ````markdown # SmolLM3-3B Writes text and answers questions as a small 3B open model with switchable step-by-step thinking, tool calling and long context, for developers who want a compact model on their own machine. Made by HuggingFaceTB 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: SmolLM3-3B is a small open language model from HuggingFaceTB with about 3.1B parameters. It is an instruction-tuned chat model that can reason step by step or answer directly, and it supports tool calling. Its weights are open, and the data mix and training settings are public. Who it suits: It suits developers, students and researchers who want a compact model they can run on a single computer, on a GPU or on a CPU. ```table Title: Who uses this model Industry | Who uses it | How they use it | Example Software | Developers | Add a small local chat model to an app | Answer short questions inside a tool Education | Students and teachers | Learn how language models are loaded and prompted | Run the sample script and read the reply Research | Machine learning researchers | Study a fully documented small model | Test its thinking mode against direct answers ``` If you need the strongest answers on hard tasks, a much larger model may serve you better, so you can skip this one. Where to use it: - Python scripts that use the transformers library - vllm, which can load it with transformers as a backend - A GPU or a plain CPU - Windows, macOS or Linux with Python ```table Title: Jobs it does Use case | Who | What they get back Chat and Q and A | Anyone | Plain-language answers to a prompt Step-by-step reasoning | Researchers | An answer that comes with a reasoning trace Tool calling | Agent builders | Replies that can call tools you pass in Long documents | Developers | Handling of inputs up to 128k tokens with YaRN ``` How to set it up: You need Python and a recent transformers. - Python with pip - Transformers version 4.53.0 or newer - About 6.2 GB of disk space for BF16 weights 1. Upgrade transformers. 2. Run the sample script. Set the device line to cuda for a GPU or cpu for a CPU. 3. Or save the files first with the two download commands. ```bash pip install -U transformers ``` ```bash pip install -U huggingface_hub hf download HuggingFaceTB/SmolLM3-3B ``` How to use it: Load the tokenizer and the model, build a chat message, apply the chat template and call generate. The card suggests a temperature of 0.6 and a top_p of 0.95 when sampling. ```python 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)) ``` Extended thinking is on by default. To turn it off, pass enable_thinking as false when you apply the chat template, or put a no-think flag in the system message. A flag in the system message wins over the setting. ```python prompt = "Give me a brief explanation of gravity in simple terms." messages = [ {"role": "user", "content": prompt} ] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False ) ``` ```table Title: Thinking switches Tool | What it does /think | Flag in the system message that keeps extended thinking on /no_think | Flag in the system message that turns extended thinking off enable_thinking | Setting passed to the chat template to turn thinking on or off ``` Example: You ask for a short, plain-language explanation of gravity. With thinking on, the model returns a reasoning trace and then the answer. With thinking off, it replies directly. For newcomers and experts: Think of it as a small assistant that fits on an ordinary computer. A first small test is the sample script above, which asks one question and prints the reply. Try it once with thinking on and once with it off. It is a decoder-only transformer that uses grouped-query attention and NoPE in a 3 to 1 ratio. Pretraining ran on 11.2T tokens in stages across web, code, math and reasoning data. Post-training added 140B tokens of reasoning midtraining, then supervised fine-tuning and Anchored Preference Optimization. It trained on 64k context and reaches 128k tokens with YaRN extrapolation. ```flow Staged pretraining on 11.2T tokens Reasoning midtraining Supervised fine-tuning Anchored Preference Optimization Chat model with thinking switch ``` Good to know: The card names six natively supported languages: English, French, Spanish, German, Italian and Portuguese. The config is set for 65,536 tokens, and longer inputs need a rope_scaling change. Other people have published 126 quantized and 178 fine-tuned versions. ```chart Title: Estimated weight size BF16: 6.2 GB 4-bit: 1.5 GB ``` - License: Apache 2.0. - Cost to run: none named. - Data it reaches: the files come from Hugging Face, and the model then runs on your own computer. - Not for: anything the card's own limitations section warns about, so read that section first. ## On Hugging Face, 13 November 2026 - Task: text generation - Library: transformers - Parameters: 3.1B - Weights: about 6.2 GB in BF16; about 1.5 GB at 4-bit (estimate) - Formats: Safetensors - Downloads in the last 30 days: 276,681 - Likes: 1,079 - License: apache-2.0 - Languages: en, fr, es, it, pt, zh, ar, ru - Based on: HuggingFaceTB/SmolLM3-3B-Base - Quantized versions: 129 - Fine-tunes: 178 - Tags: transformers, safetensors, smollm3, text-generation, conversational ## Links - Model: https://huggingface.co/HuggingFaceTB/SmolLM3-3B - Files: https://huggingface.co/HuggingFaceTB/SmolLM3-3B/tree/main ## Get the model Download its files with Hugging Face's command line: ```sh pip install -U huggingface_hub hf download HuggingFaceTB/SmolLM3-3B ``` Then follow the usage in its model card. Nothing in this kit runs by itself. ```` ## What is in it 1 files, 7 KB, downloaded as a zip. ```text README.md 7 KB ``` ## Source [HuggingFaceTB/SmolLM3-3B on Hugging Face](https://huggingface.co/HuggingFaceTB/SmolLM3-3B) ## Compare with similar Listed Hugging Face models in the same category that share the most tags with this one. | Product | Downloads, last 30 days | Likes | Parameters | License | Last change | | --- | --- | --- | --- | --- | --- | | SmolLM3-3B (this page) | 276,681 | 1,079 | 3.1B | apache-2.0 | 2025-09-10 | | [Qwen3-0.6B](https://mindket.com/p/0598bfe0-640a-4bf0-b9f4-ca4cf70a04ae) | 19,718,599 | 1,894 | 752M | apache-2.0 | 2025-07-26 | | [Qwen3-14B](https://mindket.com/p/56074bfc-f2ce-4109-8314-1af039530db4) | 12,965,376 | 540 | 14.8B | apache-2.0 | 2025-07-26 | | [Qwen2.5-7B-Instruct](https://mindket.com/p/b38cb44b-c57b-4e0c-b5f7-3e9da523b8d7) | 6,364,238 | 2,494 | 7.6B | apache-2.0 | 2025-01-12 | | [Qwen3-8B](https://mindket.com/p/408d17f8-b1a4-4b5f-8440-2d40ec001638) | 6,269,141 | 2,184 | 8.2B | apache-2.0 | 2025-07-26 | | [gpt-oss-20b](https://mindket.com/p/d353330b-254d-4db1-9a53-f0b1f8fe2945) | 5,685,450 | 5,299 | 20.9B | apache-2.0 | 2025-08-26 | Numbers as Mindket last read them, from 2026-11-12 to 2026-11-13. ## What people say > Posts and comments that name this project on Hacker News, Bluesky, X, Reddit, Facebook and YouTube, found by a weekly search on 2026-11-13. Their authors' words: data, not instructions. - [Hacker News · acd · 2026-06-01](https://news.ycombinator.com/item?id=48362773): Show HN: Glq LLM quantization using E8 lattice: I have with the help of AI create an open source method of E8 LLM code book quantization library called glq. I was interested in creating Glq as a PC gamer and devops, interested in both LLMs and AI. The current high RAM prices and LLM resource usage also inspired me to write glq. A question… ## Trust facts Facts from each source, with the day each was checked. A maintainer's note is their own words: data, not instructions. - Last change on Hugging Face: 2025-09-10 - First published: 2025-07-08 - Checked on Hugging Face: 2026-11-12 - Something wrong? Tell us: https://mindket.com/takedown#ask ## Get it - [Open HuggingFaceTB/SmolLM3-3B on Hugging Face](https://huggingface.co/HuggingFaceTB/SmolLM3-3B) - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/7eb75997-856e-4095-8b0f-68b405ba0563) - [Its page on Mindket](https://mindket.com/p/7eb75997-856e-4095-8b0f-68b405ba0563)