SmolLM model series: every generation compared
SmolLM is Hugging Face's small open language model line. Compare SmolLM, SmolLM2 and SmolLM3 by size, context length and license.
What SmolLM is and who makes it
SmolLM is a line of small language models from Hugging Face, released by its Smol Models research team. The point of the line is the opposite of scale: models in the hundreds of millions to a few billion parameters that run on one modest GPU, a laptop or a phone, with the training data, recipes and intermediate checkpoints published. The Hugging Face repository for the line lists the code, data and training configurations alongside the weights. Every card we checked for the three generations below is Apache 2.0.
The generations in order
SmolLM (July 2024)
SmolLM was released on July 16, 2024 in three sizes: 135M, 360M and 1.7B parameters. The announcement describes training on a corpus built from three parts: Cosmopedia v2 synthetic textbooks and stories (28B tokens), deduplicated FineWeb-Edu (220B tokens) and Python-Edu code (4B tokens). The 135M and 360M models saw 600B tokens, and the 1.7B model saw 1 trillion. Context length is 2,048 tokens. The 1.7B card lists Apache 2.0.
SmolLM2 (2024 to early 2025)
SmolLM2 keeps the same three sizes: 135M, 360M and 1.7B. The SmolLM2 paper was submitted on February 4, 2025 and describes the 1.7B model trained on about 11 trillion tokens in several stages, mixing web text with math, code and instruction data. It introduced three datasets that Hugging Face published: FineMath, Stack-Edu and SmolTalk. The 1.7B Instruct card lists Apache 2.0, and describes gains over the first generation in instruction following, knowledge, reasoning and math, with uses such as text rewriting, summarization and function calling. We could not source a context length for SmolLM2 from the pages we opened, so check the specific card before relying on a number.
SmolLM3 (July 2025)
SmolLM3 was released on July 8, 2025 as a single 3B model, a step up in size from the earlier 1.7B ceiling. Per the announcement and card:
- Decoder-only transformer with grouped query attention and a NoPE layout (rotary embeddings removed from every fourth layer).
- Pretrained on 11.2 trillion tokens over three stages.
- Trained at 64K context and extended to 128K with YaRN extrapolation.
- Six natively supported languages: English, French, Spanish, German, Italian and Portuguese.
- Dual-mode reasoning: add
/thinkor/no_thinkto the system prompt, or setenable_thinking, to switch extended thinking on or off. - Apache 2.0, with a base model and an instruct model published.
The repository README presents SmolLM3 as the newest release of the line.
How they were trained is part of the draw. The SmolLM2 card says the 1.7B model was trained on 256 H100 GPUs with the nanotron framework. The SmolLM3 announcement says it was trained on 384 H100 GPUs for 24 days, with mid-training on over 100B additional tokens for long context and reasoning, and the card cites 140B reasoning tokens in the mid-training stage before supervised fine-tuning. Training and evaluation use Nanotron, Lighteval and TRL, all open tools you can reuse for your own runs.
Which generation to use today
- Smallest footprint, on-device, or very cheap fine-tuning experiments: SmolLM2 at 135M or 360M. SmolLM3 has no size below 3B.
- A small general assistant with long inputs: SmolLM3, the only generation with a sourced long context (up to 128K with YaRN).
- Reasoning on and off from one checkpoint: SmolLM3, through its think and no-think flags.
- Multilingual use in the six listed European languages: SmolLM3. The earlier cards we opened did not list multilingual support.
- Reproducing a small-model training run: any of the three, since the data and recipes are published, but SmolLM3 shows the most recent end-to-end recipe.
SmolLM v1 is mostly of historical interest now: its 2,048-token context is short, and the later generations cover its sizes. Claims about how these models compare with Qwen or Llama come from Hugging Face's own pages, so test on your own data before choosing.
Running it
At 3B parameters and below, these models fit comfortably on small GPUs. You can run them on Aquanode GPUs and use each generation hub for memory estimates.
SmolLM generations
Every SmolLM generation we track, oldest first. Each hub lists all of its models with VRAM at native, FP8 and INT4 precision.
| Hub | Models | Sizes | Smallest native VRAM | Cheapest live fit for it | Est. $/hr |
|---|---|---|---|---|---|
| SmolLM-1 | 5 | 135M to 1.7B | 0.3 GB | RTX 3070 | $0.088/hr |
| SmolLM2 | 6 | 135M to 1.7B | 0.3 GB | RTX 3070 | $0.088/hr |
| SmolLM3 | 3 | 3.1B | 6.9 GB | RTX 3070 | $0.088/hr |
Smallest native VRAM is the lowest requirement among the publisher's own checkpoints in the hub, at the precision they are published in; see the methodology.
Run and fine-tune SmolLM
Per-generation guides: engine support, chat template and context flags, then fine-tune memory for LoRA and QLoRA.
- SmolLM3: how to run, fine-tune
SmolLM compared with other models
Side-by-side VRAM, context length and license for models of a similar size.
- Qwen2.5-3B-Instruct vs SmolLM3-3B-Base
- Qwen3-0.6B vs SmolLM2-135M
- Qwen2.5-3B-Instruct vs SmolLM3-3B
- Qwen3-4B vs SmolLM3-3B-Base
- Qwen2.5-1.5B-Instruct vs SmolLM2-135M
- Qwen3-4B vs SmolLM3-3B
- Qwen3-4B-Instruct-2507 vs SmolLM3-3B-Base
- Llama-3.2-1B-Instruct vs SmolLM2-135M
- Qwen3-4B-Instruct-2507 vs SmolLM3-3B
- Qwen2.5-0.5B-Instruct vs SmolLM2-135M
- Qwen3-4B-Base vs SmolLM3-3B-Base
- Qwen2.5-0.5B-Instruct vs SmolLM2-135M-Instruct
- Qwen3-4B-Base vs SmolLM3-3B
- Llama-3.2-3B-Instruct vs SmolLM3-3B-Base
- Qwen3-1.7B vs SmolLM2-135M
- Llama-3.2-3B-Instruct vs SmolLM3-3B
Best models by task
SmolLM models appear on these ranked task pages.
Sources
- https://huggingface.co/blog/smollm
- https://huggingface.co/HuggingFaceTB/SmolLM-1.7B
- https://arxiv.org/abs/2502.02737
- https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct
- https://huggingface.co/blog/smollm3
- https://huggingface.co/HuggingFaceTB/SmolLM3-3B
- https://github.com/huggingface/smollm
Facts in the text above were read from these pages and are the publisher's own statements, not benchmarks run by Aquanode. Last reviewed 2026-10-07.
Related on Aquanode: GPU requirements for SmolLM-135M-Instruct-FP32, GPU requirements for SmolLM-135M-Instruct, and GPU requirements for SmolLM-360M-Instruct.