AI Potluck
Model components / Base / pretrained models

SmolLM3

Hugging Face

Hugging Face's fully-open 3B small-model line (base + Instruct), released Jul 2025. Trained on ~11T tokens with a 128K context window and 6-language multilingual support; dual-mode reasoning. The full engineering blueprint (architecture, data mixtures, training and post-training recipe) is published. Supersedes SmolLM2.

Public data mixtures and training recipe. Outperforms Llama-3.2-3B and Qwen2.5-3B and stays competitive with 4B Qwen3/Gemma3; light enough to run on-device. Hugging Face calls the release an "engineering blueprint" and the pretraining used public datasets throughout, which is what puts a 3B model in the same openness tier as OLMo and Pythia. Verified 2026-08-13 via the SmolLM3 release post, the SmolLM3-3B model card and the huggingface/smollm repository.

Openness

5 high confidence
5.0
weights
open(Apache-2.0)
data
open(public data mixtures, ~11T tokens)
code
open(training+post-training recipe published)
license
Apache-2.0(OSI)

Fully-open small-model line: Apache-2.0 weights with the complete engineering blueprint (architecture, data mixtures, training and post-training recipe) published. The release post publishes "the complete recipe ... architecture details, exact data mixtures ... and the methodology" and describes the 11T-token run as built using public datasets and training frameworks; the card is apache-2.0 and ungated.

Adoption

3 medium confidence
3.0

849,030 downloads in the trailing 30 days for the single declared artifact HuggingFaceTB/SmolLM3-3B, which bands at level 3 (100K-1M) on the model adoption scale, near the top of that band. Strong in on-device and edge use, and niche against the large open-weight families.

Capability

3 high confidence
3.0

Strong intelligence-per-parameter for a 3B model, but small in absolute capability next to larger open and open-weight models. The comparison the score rests on is the one in the release post: SmolLM3 outperforms Llama-3.2-3B and Qwen2.5-3B and is competitive with 4B Qwen3 and Gemma 3 - a comparison against other small models rather than against the frontier, which is what a 3 encodes.

  • https://huggingface.co/blog/smollm3 recorded 2026-08-13

    "Our 3B model outperforms Llama-3.2-3B and Qwen2.5-3B while staying competitive with larger 4B alternatives (Qwen3 & Gemma3)"

Verified 2026-08-13