AI Potluck
Contribute
Back to Gap Map Model components / Base / pretrained models

SmolLM3

Hugging Face
open source / Overall score: 3.0

SmolLM3 is Hugging Face's fully open 3B small language model, released in base and instruct form in July 2025. Trained on about 11 trillion tokens, it supports a 128K context window, six languages and dual-mode reasoning. The full engineering blueprint, covering architecture, data mixtures and the training and post-training recipe, is published, and the line supersedes SmolLM2.

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)
checkpoints
open(SmolLM3-3B-checkpoints, every 40k steps)
license
Apache-2.0(OSI)

SmolLM3 publishes Apache-2.0 weights alongside a complete engineering blueprint: architecture details, exact data mixtures and training methodology for its 11-trillion-token run, built entirely from public datasets and training frameworks. The Hugging Face card is Apache-2.0 licensed and ungated.

  • https://api.github.com/repos/huggingface/smollm recorded 2026-08-13

    `private: false`, `archived: false`, license `Apache-2.0`, 3,868 stars, description "Everything about the SmolLM and SmolVLM family of models".

  • https://cdn.jsdelivr.net/gh/huggingface/smollm@main/text/pretraining/ recorded 2026-09-29

    jsDelivr listing of the huggingface/smollm repository at main shows the pretraining tree "/gh/huggingface/smollm@main/text/pretraining/smollm3/", "/gh/huggingface/smollm@main/text/pretraining/launch.slurm" and "/gh/huggingface/smollm@main/text/pretraining/continual-pretraining/", so the SmolLM3 pretraining code is public.

  • https://huggingface.co/api/models/HuggingFaceTB/SmolLM3-3B recorded 2026-09-29

    The Hub API record for the instruct checkpoint carries the tags "license:apache-2.0" and "base_model:finetune:HuggingFaceTB/SmolLM3-3B-Base"; its gated and private flags are both false, so the weights download without an access request.

  • https://huggingface.co/api/models/HuggingFaceTB/SmolLM3-3B-checkpoints recorded 2026-09-29

    The Hub API record for "HuggingFaceTB/SmolLM3-3B-checkpoints" carries the tag "license:apache-2.0" and is neither gated nor private; the model card says it holds the intermediate training checkpoints, including mid-training and SFT.

  • https://huggingface.co/api/models/HuggingFaceTB/SmolLM3-3B-checkpoints/refs recorded 2026-09-29

    Hugging Face refs API for HuggingFaceTB/SmolLM3-3B-checkpoints lists 133 branches, from "stage1-step-320000" and "stage1-step-2600000" through "lc-32k-to-64k-step-20000" and "it-mid-training".

  • https://huggingface.co/blog/smollm3 recorded 2026-09-29

    The release post describes the published recipe - "It includes architecture details, exact data mixtures showing how we progressively boost performance across domains in a three-stage pretraining approach" - and says "we're sharing exactly how we built it using public datasets and training frameworks".

  • https://huggingface.co/collections/HuggingFaceTB/smollm3-pretraining-datasets-685a7353fdc01aecde51b1d9 recorded 2026-09-29

    The SmolLM3 pretraining datasets collection lists the public corpora used, including "mlfoundations/dclm-baseline-1.0", "HuggingFaceTB/finemath", "bigcode/the-stack-v2" and "HuggingFaceTB/stackexchange_2025_md".

  • https://huggingface.co/HuggingFaceTB/SmolLM3-3B recorded 2026-09-29

    The model page says "SmolLM3 is a fully open model that offers strong performance at the 3B–4B scale" and lists "open weights + full training details including public data mixture and training configs"; its License section reads "Apache 2.0". Searched the page for agree to share and access request: none.

  • https://huggingface.co/HuggingFaceTB/SmolLM3-3B-checkpoints/raw/main/README.md recorded 2026-09-29

    SmolLM3 checkpoints card, front matter "license: apache-2.0": "We are releasing intermediate checkpoints of SmolLM3 to enable further research", "We release checkpoints every 40,000 steps, which equals 94.4B tokens", with "Stage 1:** Steps 0 to 3,450,000 (86 checkpoints)", and it lists "Pretraining scripts (nanotron)" in the linked GitHub repo.

  • https://huggingface.co/HuggingFaceTB/SmolLM3-3B/raw/main/README.md recorded 2026-09-29

    The raw model card front matter declares "license: apache-2.0", and its Training section says "The datasets used for pretraining can be found in this" linked collection and "The training and evaluation configs and code can be found in the" huggingface/smollm repository, trained with nanotron.

Adoption

3 medium confidence
3.0

Adoption is measured as Hugging Face downloads for the single declared checkpoint, SmolLM3-3B. It sees real on-device and edge use, but stays a niche choice next to the large open-weight families.

Capability

3 high confidence
3.0

SmolLM3-3B outperforms Llama-3.2-3B and Qwen2.5-3B and stays competitive with the larger 4B Qwen3 and Gemma 3. That comparison is against other small models rather than the broader frontier.

Verified 2026-09-27