HuggingFaceTB/SmolLM3-3B-Base

SmolLM3

Table of Contents

1. 1.Model Summary
2. 2.How to use
3. 3.Evaluation
4. 4.Training
5. 5.Limitations
6. 6.License

Model Summary

SmolLM3 is a 3B parameter language model designed to push the boundaries of small models. It supports 6 languages, advanced reasoning and long context. SmolLM3 is a fully open model that offers strong performance at the 3B–4B scale.

SmolLM3-3B-Base is the base model after pretraining, you can find the instruct model at SmolLM3-3B.

The model is a decoder-only transformer using GQA and NoPE, it was pretrained on 11.2T tokens with a staged curriculum of web, code, math and reasoning data. Post-training included midtraining on 140B reasoning tokens followed by supervised fine-tuning and alignment via Anchored Preference Optimization (APO).

Key features

• —Instruct model optimized for hybrid reasoning
• —Fully open model: open weights + full training details including public data mixture and training configs
• —Long context: Trained on 64k context and suppots up to 128k tokens using YARN extrapolation
• —Multilingual: 6 natively supported (English, French, Spanish, German, Italian, and Portuguese)

For more details refer to our blog post: https://hf.co/blog/smollm3

How to use

The modeling code for SmolLM3 is available in transformers v4.53.0, so make sure to upgrade your transformers version. You can also load the model with the latest vllm which uses transformers as a backend.

pip install -U transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "HuggingFaceTB/SmolLM3-3B"
device = "cuda" # for GPU usage or "cpu" for CPU usage
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
# for multiple GPUs install accelerate and do `model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto")`
model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)
inputs = tokenizer.encode("Gravity is", return_tensors="pt").to(device)
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

For local inference, you can use llama.cpp, ONNX, MLX and MLC. You can find quantized checkpoints in this collection (https://huggingface.co/collections/HuggingFaceTB/smollm3-686d33c1fdffe8e635317e23).

Long context processing

The current config.json is set for context length up to 65,536 tokens. To handle longer inputs (128k or 256k), we utilize YaRN you can change the max_position_embeddings and rope_scaling` to:

{
  ...,
  "rope_scaling": {
    "factor": 2.0, #2x65536=131 072 
    "original_max_position_embeddings": 65536,
    "type": "yarn"
  }
}

Evaluation

In this section, we report the evaluation results of SmolLM3 model. All evaluations are zero-shot unless stated otherwise, and we use lighteval to run them.

We highlight the best score in bold and underline the second-best score.

Base Pre-Trained Model

English benchmarks

Note: All evaluations are zero-shot unless stated otherwise. For Ruler 64k evaluation, we apply YaRN to the Qwen models with 32k context to extrapolate the context length.

Multilingual benchmarks

The model has also been trained on Arabic (standard), Chinese and Russian data, but has seen fewer tokens in these languages compared to the 6 above. We report the performance on these langages for information. | Category | Metric | SmolLM3 3B Base | Qwen2.5-3B | Llama3.2 3B | Qwen3 1.7B Base | Qwen3 4B Base | |---------|--------|---------------------|------------|--------------|------------------|---------------| | Other supported languages | | | | | | | | | Arabic| Belebele | 40.22 | 44.22 | <u>45.33</u> | 42.33 | 51.78 | | | Global MMLU (CF) | 28.57 | 28.81 | 27.67 | <u>29.37</u> | 31.85 | | | Flores-200 (5-shot) | <u>40.22</u> | 39.44 | 44.43 | 35.82 | 39.76 | | Chinese| Belebele | 43.78 | 44.56 | <u>49.56</u> | 48.78 | 53.22 | | | Global MMLU (CF) | 36.16 | 33.79 | <u>39.57</u> | 38.56 | 44.55 | | | Flores-200 (5-shot) | 29.17 | 33.21 | 31.89 | 25.70 | <u>32.50</u> | | Russian| Belebele | <u>47.44</u> | 45.89 | <u>47.44</u> | 45.22 | 51.44 | | | Global MMLU (CF) | <u>36.51</u> | 32.47 | 34.52 | 34.83 | 38.80 | | | Flores-200 (5-shot) | 47.13 | 48.74 | 50.74 | <u>54.70</u> | 60.53 |

Instruction Model

No Extended Thinking

Evaluation results of non reasoning models and reasoning models in no thinking mode. We highlight the best and second-best scores in bold. | Category | Metric | SmoLLM3-3B | Qwen2.5-3B | Llama3.1-3B | Qwen3-1.7B | Qwen3-4B | |---------|--------|------------|------------|-------------|------------|----------| | High school math competition | AIME 2025 | <u>9.3</u> | 2.9 | 0.3 | 8.0 | 17.1 | | Math problem-solving | GSM-Plus | 72.8 | <u>74.1</u> | 59.2 | 68.3 | 82.1 | | Competitive programming | LiveCodeBench v4 | <u>15.2</u> | 10.5 | 3.4 | 15.0 | 24.9 | | Graduate-level reasoning | GPQA Diamond | <u>35.7</u> | 32.2 | 29.4 | 31.8 | 44.4 | | Instruction following | IFEval | 76.7 | 65.6 | 71.6 | <u>74.0</u> | 68.9 | | Alignment | MixEval Hard | 26.9 | <u>27.6</u> | 24.9 | 24.3 | 31.6 | | Tool Calling | BFCL| <u>92.3</u> | - | <u>92.3</u> | 89.5 | 95.0 | | Multilingual Q&A | Global MMLU | <u>53.5</u> | 50.54 | 46.8 | 49.5 | 65.1* |

(*): this is a tool calling finetune

Extended Thinking

Evaluation results in reasoning mode for SmolLM3 and Qwen3 models: | Category | Metric | SmoLLM3-3B | Qwen3-1.7B | Qwen3-4B | |---------|--------|------------|------------|----------| | High school math competition | AIME 2025 | <u>36.7</u> | 30.7 | 58.8 | | Math problem-solving | GSM-Plus | <u>83.4</u> | 79.4 | 88.2 | | Competitive programming | LiveCodeBench v4 | 30.0 | <u>34.4</u> | 52.9 | | Graduate-level reasoning | GPQA Diamond | <u>41.7</u> | 39.9 | 55.3 | | Instruction following | IFEval | 71.2 | <u>74.2</u> | 85.4 | | Alignment | MixEval Hard | 30.8 | <u>33.9</u> | 38.0 | | Tool Calling | BFCL | <u>88.8</u> | <u>88.8</u> | 95.5 | | Multilingual Q&A | Global MMLU | <u>64.1</u> | 62.3 | 73.3 |

Training

Model

• —Architecture: Transformer decoder
• —Pretraining tokens: 11T
• —Precision: bfloat16

Software & hardware

• —GPUs: 384 H100
• —Training Framework: nanotron
• —Data processing framework: datatrove
• —Evaluation framework: lighteval
• —Post-training Framework: TRL

Open resources

Here is an infographic with all the training details.

• —The datasets used for pretraining can be found in this collection and those used in mid-training and post-training will be released in the following weeks
• —The training and evaluation configs and code can be found in the huggingface/smollm repository.
• —The training intermediate checkpoints are available at HuggingFaceTB/SmolLM3-3B-checkpoints

Limitations

SmolLM3 can produce text on a variety of topics, but the generated content may not always be factually accurate, logically consistent, or free from biases present in the training data. These models should be used as assistive tools rather than definitive sources of information. Users should always verify important information and critically evaluate any generated content.

License

Apache 2.0

Citation

@misc{bakouch2025smollm3,
  title={{SmolLM3: smol, multilingual, long-context reasoner}},
  author={Bakouch, Elie and Ben Allal, Loubna and Lozhkov, Anton and Tazi, Nouamane and Tunstall, Lewis and Patiño, Carlos Miguel and Beeching, Edward and Roucher, Aymeric and Reedi, Aksel Joonas and Gallouédec, Quentin and Rasul, Kashif and Habib, Nathan and Fourrier, Clémentine and Kydlicek, Hynek and Penedo, Guilherme and Larcher, Hugo and Morlon, Mathieu and Srivastav, Vaibhav and Lochner, Joshua and Nguyen, Xuan-Son and Raffel, Colin and von Werra, Leandro and Wolf, Thomas},
  year={2025},
  howpublished={\url{https://huggingface.co/blog/smollm3}}
}