SmolLM3-3B-Base

HuggingFaceTB Apache 2.0 out 16 months ago

3.08B parameters, text in, text out.

Verified 2026-10-25 Rank #723 in LLMs · score 71.7 method v1 Sources 3 Downloads 30d 79.7k · #342 of 619 tracked

Can I run it?1.8GB at INT4Estimatefits a 16 GB cardmethod

Can I use it?Apache 2.0Permissive; patent grantlicence file

Do people use it?79.7kdownloads, 30 days#342 of 619 we track · below the median+2 likes, downloads −16% this week178 likesHugging Face

Worth reading

Picked by rule from 1 tracked piece: the lab's own post, the most shared, a field report, the latest.

Latest · 16 months ago

Smollm3: Smol, multilingual, long-context reasoner LLM

SmolLM3 is a fully open 3B parameter model trained on 11T tokens that outperforms Llama-3.2-3B and Qwen2.5-3B while staying competitive with larger 4B alternatives.

The model removes weight decay from embedding layers to improve training stability.
Users activate reasoning or non-reasoning modes by including the /think and /no_think flags in the system prompt.

huggingface.co 2025-07-08 388 HN pts · 79 comments 3.2k words about the SmolLM3 family family

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: 27 quantized builds.

As a service0no API provider listed
On own hardware27quantized builds
Built on130104 fine-tunes · 25 adapters · 1 merges
In apps0Hugging Face Spaces

Will it fit?

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

Fits a 16 GB card at INT4

About 1.8 GB at INT4, 3.7 GB at FP8, 7.4 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.

1 dated article and 0 repo commits so far

Teal is the weights (2025-06-19), amber triangles are 1 dated article (each links to it).

What people built on it

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

157 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 mradermacher/SmolLM3-3B-Base-i1-GGUF for llama.cpp, Ollama, LM Studio 13.1k downloads MLX mlx-community/SmolLM3-3B-Base-4bit for Apple silicon 825 downloads

All variants →

Reference

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

Use-case evidence

1 use case

What it is good at 1 use case

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.

Run on a laptop or phonestructuraldecision-model estimate p=0.98 name: 3B fits: single-gpu

Facts

13 fields

Listed in LLMs

Facts

SmolLM3-3B-Base has 3.08B parameters. (safetensors metadata, observed —)
License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed —)
Created 2025-06-19, last modified 2025-08-14. (HF API, observed —)

Pipeline tag	text-generation
Library	transformers
Architectures	SmolLM3ForCausalLM
Model type	smollm3
Params by dtype	BF16: 3,075,098,624
Size band	M
MoE	no
License	Apache 2.0 (apache-2.0)
Hub snapshot	2026-10-25

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for SmolLM3-3B-Base: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	7.4 GB · Single GPU (<24 GB)
FP8 (1 byte/param)	3.7 GB · Edge (<4 GB)
INT4 (0.5 bytes/param)	1.8 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/SmolLM3 157 derivatives on the Hub

finetune: 104quantized: 27adapter: 25merge: 1

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
HuggingFaceTB/SmolLM3-3B-checkpoints	finetune	—	—	—	apache-2.0	HuggingFaceTB	25	101.9k
mradermacher/SmolLM3-3B-Base-i1-GGUF	quantized	GGUF	—	—	apache-2.0	mradermacher	0	13.1k
mradermacher/SmolLM3-3B-Base-GGUF	quantized	GGUF	—	—	apache-2.0	mradermacher	1	7.6k
DevQuasar/HuggingFaceTB.SmolLM3-3B-Base-GGUF	quantized	GGUF	—	—	—	DevQuasar	2	3.8k
unsloth/SmolLM3-3B-Base	finetune	—	3.08B	M	apache-2.0	unsloth	6	2.5k
achavan1211/SmolLM3-Custom-SFT	finetune	—	3.08B	M	—	achavan1211	0	2k
HuggingFaceTB/SmolLM3-3B-GSM8K-SFT	finetune	—	3.08B	M	apache-2.0	HuggingFaceTB	2	1.9k
mzoelfakar/Al-Khwarizmi-3B	finetune	—	3.08B	M	apache-2.0	mzoelfakar	1	1.7k
BytedTsinghua-SIA/Open-MOPD-SmolLM3-3B-MixSFT	finetune	—	3.34B	M	apache-2.0	BytedTsinghua-SIA	1	1.6k
ryjava432/lsat-smollm3-sft	finetune	—	3.08B	M	—	ryjava432	0	1.3k
Sathman/Meditation-Agent-SmolLM3-3B-v2-GGUF	quantized	GGUF	—	—	apache-2.0	Sathman	0	1.2k
Sathman/Meditation-Agent-SmolLM3-3B-GGUF	adapter	GGUF	—	—	apache-2.0	Sathman	0	1.2k
unsloth/SmolLM3-3B-Base-unsloth-bnb-4bit	quantized	-bnb	3.15B	M	apache-2.0	unsloth	2	1.1k
unsloth/SmolLM3-3B-Base-bnb-4bit	quantized	-bnb	3.16B	M	apache-2.0	unsloth	0	1k
professorf/SmolLM3-3B-gguf	quantized	gguf	—	—	apache-2.0	professorf	1	955
mlx-community/SmolLM3-3B-Base-4bit	quantized	-4bit	3.08B	M	apache-2.0	mlx-community	3	825
qiuyu8290/SmolLM3-Custom-SFT2	finetune	—	3.08B	M	—	qiuyu8290	0	795
sidgenai/smollm3-sft	finetune	—	3.08B	M	—	sidgenai	0	575
paul-stansifer/qw3-smollm3-3b-1x2e-4	quantized	GGUF	—	—	apache-2.0	paul-stansifer	0	507
bnb-community/SmolLM3-3B-Base-bnb-4bit	quantized	-bnb	3.12B	M	apache-2.0	bnb-community	0	506
liodon-ai/SmolLM3-3B-Base-ONNX	quantized	—	—	—	other	liodon-ai	0	483
espressovi/BODHI-smollm-3-maze-3b-distil	finetune	—	3.34B	M	mit	espressovi	0	480
espressovi/BODHI-smollm-3-maze-3b-rlvr	finetune	—	3.34B	M	mit	espressovi	0	448
marcelovidigal/smollm3-3b-sft-finetuned-jobs-v2	finetune	—	3.08B	M	—	marcelovidigal	0	394
4rc4n4/chrysopoeia-smollm3	adapter	GGUF	—	—	apache-2.0	4rc4n4	0	370
Cr3373/SmolLM3-3B	finetune	—	3.08B	M	apache-2.0	Cr3373	0	350
mlx-community/SmolLM3-3B-Base-bf16	finetune	-bf16	3.08B	M	apache-2.0	mlx-community	0	347
Lerelou/SmoLlm3python-3B_GGUF	adapter	GGUF	—	—	apache-2.0	Lerelou	1	290
srhm-ca/SmolLM3-3B-Base-Q4_K_M-GGUF	quantized	GGUF	—	—	apache-2.0	srhm-ca	1	285
AINovice2005/SmolLM3-3B-smashed	finetune	—	—	—	apache-2.0	AINovice2005	0	333
ShayonSarker/SmolLM3-3B-Q4_K_M-GGUF	quantized	GGUF	—	—	apache-2.0	ShayonSarker	0	322
MassivDash/SmolLM3-3B-heretic	finetune	—	3.08B	M	apache-2.0	MassivDash	0	317
Sathman/Meditation-Agent-Krishnamurti-SmolLM3-3B-v2-GGUF	quantized	GGUF	—	—	apache-2.0	Sathman	0	297
Sathman/Meditation-Agent-Watts-SmolLM3-3B-v2-GGUF	quantized	GGUF	—	—	apache-2.0	Sathman	0	272
Kibalama/smollm3-sft-01	finetune	—	3.08B	M	—	Kibalama	0	271
BigDok/Affine-5GE2VdEcMf9HTkswp4Ht579tFDVy1UByYDk9dvG4nZRnHkEs	finetune	—	3.08B	M	apache-2.0	BigDok	0	268
Sathman/Meditation-Agent-Osho-SmolLM3-3B-v2-GGUF	quantized	GGUF	—	—	apache-2.0	Sathman	0	267
Lerelou/BrainsV2.5_3B.Q4_K_M-GGUF	quantized	GGUF	—	—	apache-2.0	Lerelou	1	213
mlx-community/SmolLM3-3B-Base-8bit	quantized	-8bit	3.08B	M	apache-2.0	mlx-community	0	256
Sathman/Meditation-Agent-TNH-SmolLM3-3B-v2-GGUF	quantized	GGUF	—	—	apache-2.0	Sathman	0	254

Articles and coverage

1 piece

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.

Official1

Smollm3: Smol, multilingual, long-context reasoner LLM

huggingface.co2025-07-08officialauthorityHN front pageabout the SmolLM3 familyHN 388 pts · 79 comments3,203 words

SmolLM3 is a fully open 3B parameter model trained on 11T tokens that outperforms Llama-3.2-3B and Qwen2.5-3B while staying competitive with larger 4B alternatives. The model utilizes Grouped Query Attention with 4 groups and was trained on 384 H100 GPUs for 24 days using a global batch size of 2.36M tokens. It supports multilingual input for six languages and handles long contexts up to 128k tokens via NoPE and YaRN extrapolation.written, reviewed, cited

The model removes weight decay from embedding layers to improve training stability.
Users activate reasoning or non-reasoning modes by including the /think and /no_think flags in the system prompt.

Similar models

3 models

Similar models

Llama-3.2-3B-Instruct · 3.21B
same size band (M), same task (text-generation), more providers
Qwen2.5-3B-Instruct · 3.09B
same size band (M), same task (text-generation), more providers
Llama-3.2-3B · 3.21B
same size band (M), same task (text-generation), more providers

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-23 20:12 UTC · Last verified 2026-10-25

Follow SmolLM3-3B-Base

Watch this model

/feeds/models/HuggingFaceTB--SmolLM3-3B-Base.xml(RSS 2.0) · every change to SmolLM3-3B-Base
/feeds/families/HuggingFaceTB--SmolLM3.xml(RSS 2.0) · every change across the SmolLM3 family

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