What is SmolLM3-3B?
SmolLM3-3B is a 3B parameter language model from Hugging Face, built to push what a small model can do. It reasons in two modes: extended thinking is on by default and writes a trace before the answer, and a /no_think flag switches it off for fast direct replies. It is also fully open in a way few releases are: open weights plus the complete training details, including the public data mixture and the training configs. Hugging Face released it on July 8, 2025 under Apache 2.0, and at this size it runs on nearly any machine, no dedicated GPU required.
| Specification | SmolLM3-3B |
|---|---|
| Total parameters | 3.08B |
| Architecture | Decoder-only transformer, GQA and NoPE (3:1 ratio) |
| Pretraining | 11.2T tokens, staged web, code, math and reasoning curriculum |
| Context window | 64K trained, up to 128K tokens with YaRN |
| Languages | English, French, Spanish, German, Italian, Portuguese |
| Reasoning | Dual mode, thinking on by default, /think and /no_think flags |
| Tool calling | XML tool calls or Python-style code calls |
| Release date | July 8, 2025 |
| License | Apache 2.0 |
Post-training is where the reasoning comes from: midtraining on 140B reasoning tokens, then supervised fine-tuning and alignment with Anchored Preference Optimization. Tool calling works in two formats, JSON blobs inside XML tags or Python-style function calls in a code snippet. Six languages are supported natively; the model has also seen Arabic, Chinese and Russian data, though fewer tokens of each, so treat those as secondary. The shipped config caps context at 65,536 tokens, and YaRN rope scaling takes it to 128K. Hugging Face recommends temperature 0.6 and top_p 0.95 for sampling, and lists llama.cpp, ONNX, MLX, MLC and ExecuTorch for local inference.
SmolLM3-3B benchmarks
Hugging Face's launch numbers, from the model card, compare SmolLM3-3B with extended thinking off against Qwen2.5-3B, Llama3.1-3B, Qwen3-1.7B and Qwen3-4B:
| Benchmark | SmolLM3-3B | Qwen2.5-3B | Llama3.1-3B | Qwen3-1.7B | Qwen3-4B |
|---|---|---|---|---|---|
AIME 2025 Competition math | 9.3 | 2.9 | 0.3 | 8.0 | 17.1 |
GSM-Plus Math problems | 72.8 | 74.1 | 59.2 | 68.3 | 82.1 |
LiveCodeBench v4 Competitive coding | 15.2 | 10.5 | 3.4 | 15.0 | 24.9 |
GPQA Diamond Expert science | 35.7 | 32.2 | 29.4 | 31.8 | 44.4 |
IFEval Instruction following | 76.7 | 65.6 | 71.6 | 74.0 | 68.9 |
Global MMLU Multilingual knowledge | 53.5 | 50.54 | 46.8 | 49.5 | 65.1 |
The larger Qwen3-4B takes most rows. Inside the 3B class the picture is different: SmolLM3-3B wins instruction following outright and leads on AIME, LiveCodeBench and GPQA Diamond, while Qwen2.5-3B keeps a small edge on GSM-Plus. Switching extended thinking on moves the math and science scores sharply, with AIME 2025 climbing from 9.3 to 36.7 and GPQA Diamond from 35.7 to 41.7.
SmolLM3-3B hardware requirements
The system requirement to check is memory, and at 3B parameters almost every tier clears it. The builds and file sizes below are from the unsloth/SmolLM3-3B-GGUF repo.
| Memory | Build to pick | File size |
|---|---|---|
| 2 GB | UD-IQ2_M | 1.15 GB |
| 4 GB | Q4_K_M | 1.92 GB |
| 8 GB | Q8_0 | 3.28 GB |
| 16 GB and up | BF16 | 6.16 GB |
Neighbouring files differ by a few hundred megabytes, so when two builds both fit, take the larger one. Even the full BF16 file is only 6.16 GB, so most machines can skip the low-bit quants entirely. If the format is new to you, start with what GGUF is.
How to run SmolLM3-3B in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for SmolLM3-3B in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
For the rest of the family, see every SmolLM model you can run locally, or the pocket-sized sibling SmolLM2-135M-Instruct.
SmolLM3-3B license
SmolLM3-3B is released under Apache 2.0. That permits commercial use, modification and redistribution with no royalties, so you can build products on top of the model and run it on your own hardware without a usage fee. Hugging Face adds the usual caveat: outputs can be inaccurate or biased, so treat the model as an assistive tool rather than a source of record.
