# SmolLM3 — weights, license and what is actually open

A compact multilingual reasoning model whose engineering blueprint, mixtures, training frameworks, and alignment process are published.

Canonical: https://brightaifuture.com/open-models/smollm
Format: model-family
Source publication: Not established
Bright publication: 2026-09-19
Substantive update: None recorded
Evidence and review: AI-assisted primary-source check on 2026-09-19; no independent model reproduction or license certification. No independent capability study is attached; the release’s reproducibility artifacts make such study practical.

## Documented release

SmolLM3-3B

Organization: Hugging Face

SmolLM3 release: 2025-07-08 (day precision). Source: https://huggingface.co/blog/smollm3

Parameters: 3B

Architecture: Dense decoder transformer with grouped-query attention and alternating NoPE layers.

Modalities: text

Context: 128K tokens

## License and commercial use

Apache-2.0

Permitted by Apache-2.0; constituent training datasets have their own terms.

## What is open

Weights: available. Base and instruct/reasoning weights are public. Source: https://huggingface.co/blog/smollm3

Architecture: available. Architecture and ablations are described in detail. Source: https://huggingface.co/blog/smollm3

Inference code: available. Transformers-compatible inference and prompting instructions are public. Source: https://huggingface.co/blog/smollm3

Training code: available. Nanotron, Datatrove, and Lighteval code and configurations are linked. Source: https://huggingface.co/blog/smollm3

Training recipe: available. The three-stage data mixture and training blueprint are published. Source: https://huggingface.co/blog/smollm3

Data information: available. Sources and mixture percentages are disclosed by training stage. Source: https://huggingface.co/blog/smollm3

Training data: partial. The named public datasets are accessible, but the release is not a single immutable corpus bundle. Source: https://huggingface.co/blog/smollm3

Evaluation: available. Evaluation details and Lighteval tooling are published. Source: https://huggingface.co/blog/smollm3

Commercial use: available. The model is Apache-2.0 licensed. Source: https://huggingface.co/blog/smollm3

## Uses and strengths described in sources

Small footprint

Transparent recipe

Long context

Dual reasoning mode

## Hardware and quantization

The base parameter size makes consumer deployment practical when quantized; no single official minimum is asserted.

Community GGUF, MLX, ONNX, and bitsandbytes variants exist

## Independent evidence

No independent capability study is attached; the release’s reproducibility artifacts make such study practical.

## Limitations

Text-only

Six explicitly supported languages

Smaller capacity than frontier-scale models

## Provenance and history

{}

## Original sources

- [SmolLM3: smol, multilingual, long-context reasoner](https://huggingface.co/blog/smollm3)
- [SmolLM3-3B model card](https://huggingface.co/HuggingFaceTB/SmolLM3-3B)

## Continue exploring

- [Open Models](https://brightaifuture.com/open-models)
- [Open intelligence thread](https://brightaifuture.com/threads/open)