SmolLM2-135M

HuggingFaceTB Apache 2.0 out 2 years ago

135M parameters, text in, text out, 8K tokens of context.

Verified 2026-10-25 Rank #190 in LLMs · score 89.7 method v1 Sources 3 Downloads 30d 1.9M · #65 of 619 tracked

Changed this week: 3 new quantized builds on the Hub (78 in all).2026-10-24

Can I run it?0.1GB at INT4Estimatefits a 16 GB cardmethod

Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24

Do people use it?1.9Mdownloads, 30 days#65 of 619 we track · top 11%+10 likes, downloads +13% this week261 likesHugging Face2026-10-24

Worth reading

Picked by rule from 14 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 110 HN pts

SmolLM2

SmolLM2 is a new family of compact language models from Hugging Face researchers available in three sizes ranging from 135M to 1.7B parameters.

simonwillison.net 2024-11-02 110 HN pts · 47 comments 253 words about the SmolLM2 family family

Latest · 21 months ago

Using pip to install a Large Language Model that's under 100MB

Simon Willison released llm-smollm2, a pip-installable plugin bundling the SmolLM2-135M-Instruct.Q4_1 model.

The author used Gemini 2.0 Pro to analyze the llama-cpp-python codebase and find logging suppression methods.
The 94MB version is explicitly described as not being useful for any serious tasks.

simonwillison.net 2025-02-07 24 HN pts · 5 comments 1.2k words about the SmolLM2 family family

14 best of 14 tracked

The 2 picks above, plus these 12. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

PaperSmolLM2: When Smol Goes Big — Data-Centric Training of a Small Language Modelarxiv.org · undated
CodeThe Bunny B1: Demoing a Natrual Language to App Call Interface with SmolLM2github.com · 2024-11-05 · 14 HN pts
ArticleSmolLM2 (LLM) running in browser and WebGPUsimonwillison.net · 2024-11-29 · 2 HN pts
PaperSmolLM2: When Smol Goes Big – Data-Centric Training of a Small Language Modelarxiv.org · 2025-02-06 · 1 HN pts
ArticleSmolLM2-135M by Hugging Face Smol Models Researchsavrn.com · undated
ArticleSmolLM2-135M: Lightweight Open-Source Language Modeldev.co · undated
ArticleSmolLM2 135M — 0.135B Dense LLM | FitMyLLMfitmyllm.com · undated
DiscussionSmolLM2: The new best smol models for on-device applications (all in the open)twitter.com · 2024-11-01 · 2 HN pts
ArticleBuilding LLMs From Scratch: Part 2— SmolLM2 & Llama | by Shashank Sane | Mediummedium.com · 2026-02-27
ArticleHugging Face and Pruna AI: SmolLM2, Now 7x Smaller and 2x Fasterpruna.ai · 2025-01-06
Articlesmollm2ollama.com · 2024-11-01
ArticleSmolLM2, SmolVLM & Smol Models: Hugging Face's SLM Familyturingpost.com · 2025-02-26

Show all 12

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 78 quantized builds and 100 Spaces.

As a service0no API provider listed
On own hardware78quantized builds
Built on1,015987 fine-tunes · 26 adapters · 2 merges
In apps100+Hugging Face Spaces

Top Spaces running it: pliny-the-prompter/obliteratus ♥ 621 · embedl/hfviewer ♥ 181 · fromziro/SLM_Regression_Line ♥ 16 · stanley-00/slm-testing ♥ 9

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 16 GB card at INT4

About 0.1 GB at INT4, 0.2 GB at FP8, 0.3 GB at BF16: weights only, so leave room for the KV cache.

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2024-10-31), amber triangles are 6 dated articles (each links to it), grey ticks are repo commits.

What people built on it

1,093 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

1,093 derivatives on the Hub, most of them fine-tunes

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

GGUF prithivMLmods/SmolLM2-135M-GGUF for llama.cpp, Ollama, LM Studio 170.4k downloads GPTQ lilmeaty/SmolLM2-135M-Instruct-GPTQ for vLLM, SGLang 101.7k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

17 rows

Benchmarks 17 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (17)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
author-reported ARC (Average)	43.9	SmolLM2-135M-8k	model author	2026-10-07
author-reported ARC (Average)	37.3	SmolLM2-135M-Instruct	model author	2026-10-07
author-reported BBH (3-shot)	28.2	SmolLM2-135M-Instruct	model author	2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions	68.4	SmolLM2-135M-8k	model author	2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions	33.9	SmolLM2-135M-8k	model author	2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions	4.1	SmolLM2-135M-8k	model author	2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions	34.6	SmolLM2-135M-8k	model author	2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions	66.3	SmolLM2-135M-Instruct	model author	2026-10-07
author-reported GSM8K · Reason through hard problems	1.4	SmolLM2-135M-8k	model author	2026-10-07
author-reported GSM8K · Reason through hard problems	1.4	SmolLM2-135M-Instruct	model author	2026-10-07
author-reported HellaSwag · Chat and follow instructions	42.1	SmolLM2-135M-8k	model author	2026-10-07
author-reported HellaSwag · Chat and follow instructions	40.9	SmolLM2-135M-Instruct	model author	2026-10-07
author-reported IFEval · Chat and follow instructions	29.9	SmolLM2-135M-Instruct	model author	2026-10-07
author-reported MMLU · Reason through hard problems	31.5	SmolLM2-135M-8k	model author	2026-10-07
author-reported MMLU · Reason through hard problems	29.3	SmolLM2-135M-Instruct	model author	2026-10-07
author-reported MT-Bench · Chat and follow instructions	1.98	SmolLM2-135M-Instruct	model author	2026-10-07
author-reported Winogrande · Chat and follow instructions	51.3	SmolLM2-135M-8k	model author	2026-10-07

What the card says

1 cited sentence

Summary generated, cited

This compact language model is designed for on-device use and is released under the Apache 2.0 license.^[1]^[2]

“lightweight enough to run on-device”
license

Generated 2026-10-07 15:04 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

3 use cases

What it is good at 3 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

21 fields

Listed in LLMs

Facts

SmolLM2-135M has 135M parameters. (safetensors metadata, observed 2026-10-24)
Context length is 8,192 tokens (8K). (config.json max_position_embeddings, observed 2026-10-24)
License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed 2026-10-24)
Created 2024-10-31, last modified 2025-02-06. (HF API, observed 2026-10-24)

Pipeline tag	text-generation
Library	transformers
Architectures	LlamaForCausalLM
Model type	llama
Dtype	bfloat16
Params by dtype	BF16: 134,515,008
Size band	S
MoE	no
Context length	8,192 (config.json max_position_embeddings)
Languages	en
Papers	arXiv:2502.02737
License	Apache 2.0 (apache-2.0)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · LlamaForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · LlamaForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for SmolLM2-135M: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	0.3 GB · Edge (<4 GB)
FP8 (1 byte/param)	0.2 GB · Edge (<4 GB)
INT4 (0.5 bytes/param)	0.1 GB · Edge (<4 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

40 variants

Lineage

Original model Family: HuggingFaceTB/SmolLM2 1,093 derivatives on the Hub

quantized: 78finetune: 987merge: 2adapter: 26

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
HuggingFaceTB/SmolLM2-135M-Instruct	quantized	—	135M	S	apache-2.0	HuggingFaceTB	462	16.9M
prithivMLmods/SmolLM2-135M-GGUF	quantized	GGUF	—	—	creativeml-openrail-m	prithivMLmods	3	170.4k
lilmeaty/SmolLM2-135M-Instruct-GPTQ	quantized	GPTQ	135M	S	apache-2.0	lilmeaty	0	101.7k
unsloth/SmolLM2-135M	finetune	—	135M	S	apache-2.0	unsloth	7	51.9k
unsloth/SmolLM2-135M-bnb-4bit	quantized	-bnb	138M	S	apache-2.0	unsloth	4	17.9k
HuggingFaceTB/smollm2-135M-SFT-Only	finetune	—	135M	S	—	HuggingFaceTB	1	6.6k
onnx-community/SmolLM2-135M-ONNX	quantized	—	—	—	—	onnx-community	3	5.5k
mradermacher/SmolLM2-135M-i1-GGUF	quantized	GGUF	—	—	apache-2.0	mradermacher	0	5.4k
ggml-org/SmolLM2-135M-GGUF	quantized	GGUF	—	—	apache-2.0	ggml-org	1	4.3k
mradermacher/SmolLM2-135M-GGUF	quantized	GGUF	—	—	apache-2.0	mradermacher	1	3.9k
KrynexLabs/KrynexAI-1.4-150M-Mobile-Model	quantized	GGUF	135M	S	apache-2.0	KrynexLabs	2	3.5k
Mungert/SmolLM2-135M-Instruct-GGUF	quantized	GGUF	—	—	apache-2.0	Mungert	1	3.4k
syvai/plapre-pico	quantized	GGUF	118M	S	cc-by-4.0	syvai	10	2.7k
open-machine/SmolLM2-135M-FlashNorm	finetune	—	134M	S	apache-2.0	open-machine	3	2k
nlpguy/smolchess-v2	finetune	—	135M	S	apache-2.0	nlpguy	1	2k
hathucviet/SmolLM2-135M-sft	finetune	—	135M	S	—	hathucviet	0	1.8k
pngwn/SmolLM2-135M-Capybara-SFT	finetune	—	135M	S	—	pngwn	0	1.7k
Zyroxx66/SmolLM2-135M-FunctionCalling-100K	finetune	—	135M	S	—	Zyroxx66	0	1.7k
Emil7018/business-news-generator	finetune	—	135M	S	apache-2.0	Emil7018	1	1.6k
diegoquinteiro/SmolLM2-135M-Observable	quantized	—	—	—	apache-2.0	diegoquinteiro	0	1.5k
namanbnsl/smollm2-135m-code-cpu	finetune	—	135M	S	—	namanbnsl	1	1.3k
SebastianBodza/SmolKartoffel-135M-v0.1	finetune	—	172M	S	cc-by-nc-4.0	SebastianBodza	1	1.2k
ewre324/ewre324-Thinker-SmolLM2-135M-Instruct-Reasoning	finetune	—	135M	S	apache-2.0	ewre324	1	1.1k
gvij/SmolLM2-135M-Function-Calling	finetune	—	135M	S	apache-2.0	gvij	2	965
gugukaka/SmolLM2-FT-MyDataset	finetune	—	135M	S	—	gugukaka	0	1.1k
TobDeBer/SmolLM2-135M-Instruct-hirma-b60s-0.5	finetune	—	99.6M	S	apache-2.0	TobDeBer	0	1k
TobDeBer/SmolLM2-135M-Instruct-magic1	finetune	—	135M	S	apache-2.0	TobDeBer	0	980
QuantLLM/SmolLM2-135M-GGUF	quantized	GGUF	—	—	apache-2.0	QuantLLM	1	911
pngwn/sft-out	finetune	—	135M	S	—	pngwn	0	950
kenpeter123/small-1b-pretrain	finetune	—	1.03B	S	apache-2.0	kenpeter123	0	914
Chloemp/smollm2-135m-xlam-fullft	finetune	—	135M	S	apache-2.0	Chloemp	0	911
LMLK/SmolLM2-135M-Instruct-GGUF	quantized	GGUF	—	—	apache-2.0	LMLK	0	882
vonjack/SmolLM2-135M-Merged	merge	GGUF	135M	S	—	vonjack	0	872
ghananlpcommunity/kasawa	finetune	—	135M	S	cc-by-4.0	ghananlpcommunity	1	818
1234Frede/SmolLM2-business-news-generator	finetune	—	135M	S	apache-2.0	1234Frede	0	864
jeanlee/SmolLM2-FT-MyDataset-2026	finetune	—	135M	S	—	jeanlee	0	840
Mathildeholst/Good-generator	finetune	—	135M	S	apache-2.0	Mathildeholst	0	803
aritroroy0/SmolLM2-FT-MyDataset	finetune	—	135M	S	apache-2.0	aritroroy0	0	801
1234Frede/Bad_SmolLM2-business-news-generator	finetune	—	135M	S	apache-2.0	1234Frede	0	740
brtenorio/SmolLM2-FT-LoRA-gsm8k	finetune	—	135M	S	—	brtenorio	0	709

Repository changelog

10 commits

changed in the last 7 daysOther: 2026-10-24all changes

Repository changelog 10 commits

Date	Commit	SHA
2025-02-06	Update README.md	93efa2f
2024-11-23	Update README.md	4e53f73
2024-11-05	Update config.json	28e66ca
2024-10-31	Update README.md	acb1e6c
2024-10-31	Update README.md	615de5c
2024-10-31	Update README.md	2b7c3b9
2024-10-31	Update README.md	f102afc
2024-10-31	Upload tokenizer	ab35b5b
2024-10-31	Upload LlamaForCausalLM	d6a5589
2024-10-31	initial commit	bbf7524

Articles and coverage

14 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Analysis9

SmolLM2

simonwillison.net2024-11-02authorityHN front pageabout the SmolLM2 familyHN 110 pts · 47 comments253 words

SmolLM2 is a new family of compact language models from Hugging Face researchers available in three sizes ranging from 135M to 1.7B parameters. Trained on 11 trillion tokens using FineWeb-Edu, DCLM, and The Stack, these lightweight models are designed to run on-device and solve a wide range of tasks efficiently. The model weights are released under an Apache 2 license.written, reviewed, cited

SmolLM2: When Smol Goes Big — Data-Centric Training of a Small Language Model

arxiv.orgauthorityabout the SmolLM2 family12,445 words

SmolLM2 is a 1.7 billion parameter language model developed through data-centric training. The model was overtrained on ~11 trillion tokens using a multi-stage process that mixes web text with specialized math, code, and instruction-following data. Training required around 1e23 FLOPs, costing $250,000 USD in GPU compute. New datasets FineMath, Stack-Edu, and SmolTalk were created to address quality issues in existing sources.written, reviewed, cited

Math and code capabilities typically emerge only after extensive training on a mid-training checkpoint.
Code ablations utilized 200B tokens uniformly distributed across 15 of the most commonly used programming languages.

Using pip to install a Large Language Model that's under 100MB

simonwillison.net2025-02-07authorityabout the SmolLM2 familyHN 24 pts · 5 comments1,165 words

Simon Willison released llm-smollm2, a pip-installable plugin bundling the SmolLM2-135M-Instruct.Q4_1 model. The package is 94MB to comply with PyPI limits, utilizing the llama-cpp-python library with verbose=False and logging.CRITICAL to suppress output. While the model is not useful for serious tasks, it serves as a functional example of running a quantized LLM via pip.written, reviewed, cited

The author used Gemini 2.0 Pro to analyze the llama-cpp-python codebase and find logging suppression methods.
The 94MB version is explicitly described as not being useful for any serious tasks.

The Bunny B1: Demoing a Natrual Language to App Call Interface with SmolLM2

github.com2024-11-05authorityabout the SmolLM2 familyHN 14 pts · 0 comments265 words

The Bunny B1 demo combines SmolLM2 with Outlines to create a natural language interface for local apps on smaller devices. It requires a Rust compiler and allows users to edit functions.json to add new actions like ordering food or rides. The model defaults to the mps device on Macs but can be swapped for cuda.written, reviewed, cited

SmolLM2 (LLM) running in browser and WebGPU

simonwillison.net2024-11-29authorityabout the SmolLM2 familyHN 2 pts · 0 comments222 words

Hugging Face's SmolLM2-1.7B-Instruct runs directly in a web browser using WebGPU to perform structured text extraction. The demo relies on web-llm by MLC and its XGrammar library for JSON schema and EBNF-based guidance, requiring Chrome for execution. It loads 924.8MB of model data to convert image descriptions into GitHub issues defined by JSON schema.written, reviewed, cited

SmolLM2: When Smol Goes Big – Data-Centric Training of a Small Language Model

arxiv.org2025-02-06authorityabout the SmolLM2 familyHN 1 pts · 0 comments368 words

SmolLM2 is a 1.7 billion parameter language model that outperforms recent small LMs including Qwen2.5-1.5B and Llama3.2-1B. The model trains on ~11 trillion tokens using a multi-stage process that mixes web text with specialized math, code, and instruction-following data. New specialized datasets named FineMath, Stack-Edu, and SmolTalk were introduced to address limitations in existing data quality and size.written, reviewed, cited

SmolLM2-135M by Hugging Face Smol Models Research

savrn.com1,356 words

SmolLM2-135M: Lightweight Open-Source Language Model

dev.co938 words

SmolLM2 135M — 0.135B Dense LLM | FitMyLLM

fitmyllm.com143 words

Community1

SmolLM2: The new best smol models for on-device applications (all in the open)

twitter.com2024-11-01about the SmolLM2 familyHN 2 pts · 0 comments

All coverage tracked for this model (14) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
SmolLM2: When Smol Goes Big — Data-Centric Training of a Small Language Model	arxiv.org	—	2026-10-07	exa
SmolLM2-135M by Hugging Face Smol Models Research	savrn.com	—	2026-10-07	exa
SmolLM2-135M: Lightweight Open-Source Language Model	dev.co	—	2026-10-07	exa
SmolLM2 135M — 0.135B Dense LLM \| FitMyLLM	fitmyllm.com	—	2026-10-07	exa
Building LLMs From Scratch: Part 2— SmolLM2 & Llama \| by Shashank Sane \| Medium	medium.com	2026-02-27	2026-10-07	exa
SmolLM2, SmolVLM & Smol Models: Hugging Face's SLM Family	turingpost.com	2025-02-26	2026-10-07	exa
Using pip to install a Large Language Model that's under 100MB	simonwillison.net	2025-02-07	2026-10-07	hn
SmolLM2: When Smol Goes Big – Data-Centric Training of a Small Language Model	arxiv.org	2025-02-06	2026-10-07	hn
Hugging Face and Pruna AI: SmolLM2, Now 7x Smaller and 2x Faster	pruna.ai	2025-01-06	2026-10-07	hn
SmolLM2 (LLM) running in browser and WebGPU	simonwillison.net	2024-11-29	2026-10-07	hn
The Bunny B1: Demoing a Natrual Language to App Call Interface with SmolLM2	github.com	2024-11-05	2026-10-07	hn
SmolLM2	simonwillison.net	2024-11-02	2026-10-07	hn
smollm2	ollama.com	2024-11-01	2026-10-07	hn
SmolLM2: The new best smol models for on-device applications (all in the open)	twitter.com	2024-11-01	2026-10-07	hn

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model this model

Loubna Ben Allal, Anton Lozhkov, Elie Bakouch +19 · 2025-02-04 · PDF · HF paper page (260 upvotes)

Similar models

3 models

Similar models

gpt2 · 137M
same size band (S), same task (text-generation), permissive license
SmolLM-135M · 135M
same size band (S), same task (text-generation), permissive license
gpt-neo-125m · 150M
same size band (S), same task (text-generation), permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:36 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:04 UTC · Last verified 2026-10-25

Follow SmolLM2-135M

Watch this model

/feeds/models/HuggingFaceTB--SmolLM2-135M.xml(RSS 2.0) · every change to SmolLM2-135M
/feeds/families/HuggingFaceTB--SmolLM2.xml(RSS 2.0) · every change across the SmolLM2 family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.