gemma-3-4b-pt

google Gemma Terms of Use out 20 months ago

4.30B parameters, images and text in, text out. 1 provider serves it.

Verified 2026-10-25 Rank #330 in LLMs · score 82.1 method v1 Sources 3 Downloads 30d 250.6k · #232 of 619 tracked

Can I run it?2.6GB at INT4Estimatefits a 16 GB cardmethod

Can I use it?Gemma Terms of UseGoogle's terms + prohibited-use policy; must pass terms downstreamlicence file2026-10-24

Do people use it?250.6kdownloads, 30 days#232 of 619 we track · top 37%+0 likes, downloads +93% this week172 likesHugging Face2026-10-24

Worth reading

Picked by rule from 26 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 488 HN pts

Gemma 3 Technical Report [pdf]

storage.googleapis.com 2025-03-12 488 HN pts · 245 comments about the gemma 3 family family

Latest · 38 days ago

Open-jev: One-pass option scoring with Gemma 3 4B, similar to jev

An open-source implementation of Jev option scoring runs on local Gemma 3 4B models using MLX on Apple silicon or PyTorch on other platforms.

Frozen-feature extraction and head training workflows currently require MLX on Apple silicon.
PMI normalization requires an extra batched pass to subtract the option's unconditional log-prob.

github.com 2026-09-17 7 HN pts · 1 comments 2.1k words

25 best of 26 tracked

The 2 picks above, plus these 23. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

CodeShow HN: Gemma 3 inference in pure C++ with Metal accelerationgithub.com · 2026-07-04 · 6 HN pts
ArticleGemma 3 270M: Compact model for hyper-efficient AIdevelopers.googleblog.com · 2025-08-14 · official
ArticleGemma 3 QAT Models: Bringing AI to Consumer GPUsdevelopers.googleblog.com · 2025-04-18 · 602 HN pts · official
ArticleIntroducing Gemma 3ndevelopers.googleblog.com · 2025-06-26 · 405 HN pts · official
ArticleGemma 3n preview: Mobile-first AIdevelopers.googleblog.com · 2025-05-20 · 441 HN pts · official
ArticleGemma3 Function Callingai.google.dev · 2025-03-23 · 202 HN pts · official
ArticleFine-tune Google's Gemma 3unsloth.ai · 2025-03-19 · 226 HN pts
CodeGemma 3 270M re-implemented in pure PyTorch for local tinkeringgithub.com · 2025-08-20
CodeI have written gemma3 inference in pure Cgithub.com · 2026-01-26 · 65 HN pts
ArticleGemma3 – The current strongest model that fits on a single GPUollama.com · 2025-03-12 · 252 HN pts
ArticleGoogle calls Gemma 3 the most powerful AI model you can run on one GPUtheverge.com · 2025-03-20
ArticleGemma 3: Google's new multimodal modelsai.google.dev · 2025-03-12 · official
ArticleGemma 3: The most capable model you can run on a single GPU or TPUblog.google · 2025-03-12 · official
ArticleOwn your AI: Learn how to fine-tune Gemma 3 270M and run it on-devicedevelopers.googleblog.com · 2025-10-10 · official
ArticleGemma 3 QAT Modelssimonwillison.net · 2025-04-19
DiscussionGemma 3n Architectural Innovations – Speculation and poking around in the modelold.reddit.com · 2025-05-25
ArticleGemma 3n available in the open-source ecosystemhuggingface.co · 2025-06-26
ArticleGemma 3n: The Developer Guidesimonwillison.net · 2025-06-27
ArticleLandrecords – cheap nationwide parcel dataset standardized using gemma3landrecords.us · 2025-08-20
CodeA C++ library to efficiently run Gemma-3N across various platformgithub.com · 2025-06-11
ArticleSelectively reducing eval awareness and murder in Gemma 3 27B via steeringlesswrong.com · 2026-03-12
ArticleI Gave Up Trying to Run Gemma 3 27B on AWS P3dn.24xlargemedium.com · 2025-08-11
ArticleGoogle's Gemma 3 is an open source, multimodal single-GPU AI 128K context windowarstechnica.com · 2025-03-13

Show all 23

Where to run it

Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.

#	Cheapest live row	In /1M	Out /1M	Context	Quant	Seen
1	Featherless	—	—	—	—	Oct 25

All 1 provider →

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.

Served by 1 API provider; 26 quantized builds for local use; a big base for builders (368 fine-tunes, adapters and merges).

As a service1Featherless
On own hardware26quantized builds
Built on368348 fine-tunes · 19 adapters · 1 merges
In apps13Hugging Face Spaces

Top Spaces running it: stats-powered-ai/StatDetectLLM ♥ 9 · ariG23498/gemma3-license-plate-detection ♥ 4 · Jasmineame/detect-LLM-tokens ♥ 4 · Lunamos/flas-demo ♥ 2

Will it fit?

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

Fits a 16 GB card at INT4

About 2.6 GB at INT4, 5.2 GB at FP8, 10 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.

10 dated articles and 4 repo commits so far

Teal is the weights (2025-02-20), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.

What people built on it

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

394 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 DevQuasar-5/google.gemma-3-4b-pt-GGUF for llama.cpp, Ollama, LM Studio 10.8k downloads MLX mlx-community/gemma-3-4b-it-4bit for Apple silicon 54.6k downloads

All variants →

Reference

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

Benchmarks

21 rows

Benchmarks 21 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (21)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 AGIEval	42.1	Gemma 3 PT 4B	model author	2026-10-07
author-reported ARC Challenge · Chat and follow instructions	56.2	Gemma 3 PT 4B	model author	2026-10-07
author-reported BBH · Reason through hard problems	50.9	Gemma 3 PT 4B	model author	2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions	72.3	Gemma 3 PT 4B	model author	2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions	79.6	Gemma 3 PT 4B	model author	2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions	51.9	Gemma 3 PT 4B	model author	2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions	65.8	Gemma 3 PT 4B	model author	2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions	20	Gemma 3 PT 4B	model author	2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions	82.4	Gemma 3 PT 4B	model author	2026-10-07
author-reported DROP	60.1	Gemma 3 PT 4B	model author	2026-10-07
author-reported GPQA Diamond · Reason through hard problems	15	Gemma 3 PT 4B	model author	2026-10-07
author-reported GSM8K · Reason through hard problems	38.4	Gemma 3 PT 4B	model author	2026-10-07
author-reported Global-MMLU · Work across languages	57	Gemma 3 PT 4B	model author	2026-10-07
author-reported HellaSwag · Chat and follow instructions	77.2	Gemma 3 PT 4B	model author	2026-10-07
author-reported HumanEval · Write code with agents	36	Gemma 3 PT 4B	model author	2026-10-07
author-reported MATH-500 · Reason through hard problems	24.2	Gemma 3 PT 4B	model author	2026-10-07
author-reported MBPP · Write code with agents	46	Gemma 3 PT 4B	model author	2026-10-07
author-reported MGSM · Work across languages	34.7	Gemma 3 PT 4B	model author	2026-10-07
author-reported MMLU · Reason through hard problems	59.6	Gemma 3 PT 4B	model author	2026-10-07
author-reported MMLU-Pro · Reason through hard problems	29.2	Gemma 3 PT 4B	model author	2026-10-07
author-reported Winogrande · Chat and follow instructions	64.7	Gemma 3 PT 4B	model author	2026-10-07

Use-case evidence

7 use cases

What it is good at 7 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

18 fields

Listed in LLMs

Facts

gemma-3-4b-pt has 4.30B parameters. (safetensors metadata, observed 2026-10-24)
License: Gemma Terms of Use (friction: attribution). Google's terms + prohibited-use policy; must pass terms downstream. (license file, observed 2026-10-24)
Created 2025-02-20, last modified 2025-03-21. (HF API, observed 2026-10-24)
Access is gated on Hugging Face (manual). (HF API, observed 2026-10-24)

Pipeline tag	image-text-to-text
Library	transformers
Architectures	Gemma3ForConditionalGeneration
Model type	gemma3
Dtype	bfloat16
Params by dtype	BF16: 4,300,079,472
Size band	M
MoE	no
Gated	manual
Papers	arXiv:1905.07830, arXiv:1905.10044, arXiv:1911.11641, arXiv:1904.09728, arXiv:1705.03551, arXiv:1911.01547, arXiv:1907.10641, arXiv:1903.00161, arXiv:2009.03300, arXiv:2304.06364, arXiv:2103.03874, arXiv:2110.14168, arXiv:2311.12022, arXiv:2108.07732, arXiv:2107.03374, arXiv:2210.03057, arXiv:2106.03193, arXiv:1910.11856, arXiv:2502.12404, arXiv:2502.21228, arXiv:2404.16816, arXiv:2104.12756, arXiv:2311.16502, arXiv:2203.10244, arXiv:2404.12390, arXiv:1810.12440, arXiv:1908.02660, arXiv:2312.11805
License	Gemma Terms of Use (gemma)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Gemma3ForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Gemma3ForConditionalGeneration · 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 gemma-3-4b-pt: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	10 GB · Single GPU (<24 GB)
FP8 (1 byte/param)	5.2 GB · Single GPU (<24 GB)
INT4 (0.5 bytes/param)	2.6 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: google/gemma-3 394 derivatives on the Hub

finetune: 348quantized: 26adapter: 19merge: 1

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
google/shieldgemma-2-4b-it	finetune	—	4.30B	M	gemma	google	182	293.8k
unsloth/gemma-3-4b-pt-unsloth-bnb-4bit	quantized	-bnb	4.39B	M	gemma	unsloth	4	184.6k
mlx-community/gemma-3-4b-it-8bit	finetune	-8bit	4.97B	M	gemma	mlx-community	6	150.8k
unsloth/gemma-3-4b-pt	finetune	—	4.30B	M	gemma	unsloth	6	150.3k
twinkle-ai/gemma-3-4B-T1-it	finetune	—	3.88B	M	gemma	twinkle-ai	41	91.7k
hi-wesley/gemma3-vision-encoder	finetune	—	417M	S	gemma	hi-wesley	1	85.5k
CEIA-UFG/Gemma-3-Gaia-PT-BR-4b-it	finetune	—	4.30B	M	gemma	CEIA-UFG	159	69k
google/medgemma-4b-pt	finetune	—	4.30B	M	other	google	164	58.4k
mlx-community/gemma-3-4b-it-4bit	quantized	-4bit	4.97B	M	gemma	mlx-community	8	54.6k
KISTI-KONI/KONI-4B-instruct-20250901	finetune	—	3.88B	M	gemma	KISTI-KONI	3	50.3k
unsloth/gemma-3-4b-pt-bnb-4bit	quantized	-bnb	4.41B	M	gemma	unsloth	1	25.6k
MBZUAI-Paris/Nile-Chat-4B	finetune	—	3.88B	M	gemma	MBZUAI-Paris	20	18.1k
VIDraft/Gemma-3-R1984-4B	finetune	—	4.30B	M	gemma	VIDraft	133	8.5k
mozilla-ai/gemma-3-4b-it-llamafile	finetune	—	—	—	gemma	mozilla-ai	7	11.7k
mlx-community/gemma-3-4b-it-bf16	finetune	-bf16	4.97B	M	gemma	mlx-community	2	11.3k
DevQuasar-5/google.gemma-3-4b-pt-GGUF	quantized	GGUF	—	—	—	DevQuasar-5	2	10.8k
mradermacher/gemma-3-4b-pt-i1-GGUF	quantized	GGUF	—	—	gemma	mradermacher	1	10.3k
Mungert/Gemma-3-Gaia-PT-BR-4b-it-GGUF	quantized	GGUF	—	—	gemma	Mungert	0	8.1k
Mungert/medgemma-4b-pt-GGUF	quantized	GGUF	—	—	other	Mungert	0	7.1k
McGill-NLP/AfriqueGemma-4B	finetune	—	4.30B	M	cc-by-4.0	McGill-NLP	0	6.5k
mradermacher/gemma-3-4b-pt-GGUF	quantized	GGUF	—	—	gemma	mradermacher	0	5.9k
KISTI-KONI/KONI-4B-base-20250819	finetune	—	3.88B	M	gemma	KISTI-KONI	3	5.7k
ginipick/Gemma-3-R1984-4B	finetune	—	4.30B	M	gemma	ginipick	52	2.8k
google/gemma-3-4b-pt-qat-q4_0-gguf	quantized	q4_0	—	—	gemma	google	29	3.6k
ariG23498/gemma-3-4b-pt-object-detection	finetune	—	4.30B	M	apache-2.0	ariG23498	6	3.3k
llmvision/glimpse-v1	quantized	GGUF	4.30B	M	gemma	llmvision	9	3k
luel/gemma-3-4b-tigrinya	finetune	—	3.88B	M	gemma	luel	1	3k
tbmod/gemma-3-4b-it	finetune	—	4.30B	M	gemma	tbmod	1	1.5k
DrRiceIO7/gemma-3-4b-it-heretic-creative	finetune	—	4.30B	M	gemma	DrRiceIO7	2	1.4k
Abdusin/gemma-3-4b-est-v1	quantized	GGUF	4.30B	M	gemma	Abdusin	1	1.3k
professorf/gemma-3-4b-it-gguf	quantized	gguf	—	—	gemma	professorf	1	1.1k
manthilaffs/Gamunu-4B-Instruct-Alpha	finetune	—	4.30B	M	apache-2.0	manthilaffs	0	810
zeromodels/gemma-3-4b-pt	finetune	—	—	—	gemma	zeromodels	0	777
thaitung/gemma-3-4b-pt-Q4_K_M-GGUF	quantized	GGUF	—	—	gemma	thaitung	0	766
ngxson/gemma-3-4b-pt-Q4_0-GGUF	quantized	Q4_0	—	—	gemma	ngxson	1	669
MBZUAI-Paris/Nile-Chat-3x4B-A6B	finetune	—	9.23B	L	gemma	MBZUAI-Paris	3	560
mlx-community/gemma-3-4b-it-6bit	finetune	—	4.97B	M	gemma	mlx-community	0	591
coder3101/gemma-3-4b-it-heretic	finetune	—	4.30B	M	gemma	coder3101	3	424
munkhbayar-batkhuu/gemma-3-4b-mongolian	finetune	—	4.55B	M	gemma	munkhbayar-batkhuu	0	569
GoYM/gemma-ft	adapter	—	—	—	—	GoYM	0	506

Repository changelog

4 commits

Repository changelog 4 commits

Date	Commit	SHA
2025-03-21	Update README.md	cc012e0
2025-03-19	Adding top_k and top_p to generation_config.json	2166bd1
2025-03-12	Update README.md	158684e
2025-03-12	Super-squash branch 'main' using huggingface_hub	302f3cd

All providers

1 provider · 1 endpoint

Where to run 1 providers · 1 endpoints

Hosted endpoints observed for gemma-3-4b-pt, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting.

Provider	Model ref	Context	Quant	$/M in	$/M out	tok/s	TTFT	Tools	Status	Observed
Featherless hf	google/gemma-3-4b-pt	—	—	—	—	—	—	—	live	2026-10-25

Articles and coverage

26 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.

Official4

Gemma 3 QAT Models: Bringing AI to Consumer GPUs

developers.googleblog.com2025-04-18officialauthorityHN front pageabout the gemma 3 familyHN 602 pts · 276 comments1,077 words

Google released Gemma 3 Quantization-Aware Training (QAT) models that incorporate quantization during training to reduce memory usage while maintaining accuracy. Applying QAT on ~5,000 steps reduces the perplexity drop by 54% when quantizing down to Q4_0, allowing the 27B variant to fit in 14.1 GB of VRAM instead of 54 GB. These int4 models run efficiently on consumer hardware like the NVIDIA RTX 3090 and are available via Ollama, llama.cpp, and MLX.written, reviewed, cited

Deploying the largest models often requires lower-precision formats like FP8, potentially accepting a performance trade-off for feasibility.
The Gemmaverse offers many alternatives using Post-Training Quantization (PTQ) with contributions from members such as Bartowski, Unsloth, and GGML.

Introducing Gemma 3n

developers.googleblog.com2025-06-26officialauthorityHN front pageabout the gemma 3 familyHN 405 pts · 191 comments1,690 words

Gemma 3n is a mobile-first AI model family designed for on-device use with multimodal capabilities. The E2B and E4B versions have raw parameters of 5B and 8B but operate with as little as 2GB and 3GB of memory respectively. It supports 140 languages for text and includes a MobileNet-V5-300M vision encoder that delivers a 13x speedup with quantization, requires 46% fewer parameters, and has a 4x smaller memory footprint.written, reviewed, cited

Developers can create custom-sized models between E2B and E4B by adjusting feed forward network hidden dimensions from 8192 to 16384.
The Gemma 3n audio encoder processes clips up to 30 seconds at launch but supports arbitrarily long audios with additional training.

Gemma 3n preview: Mobile-first AI

developers.googleblog.com2025-05-20officialauthorityHN front pageabout the gemma 3 familyHN 441 pts · 163 comments904 words

Gemma3 Function Calling

ai.google.dev2025-03-23officialauthorityHN front pageabout the gemma 3 familyHN 202 pts · 32 comments2,144 words

Analysis6

Open-jev: One-pass option scoring with Gemma 3 4B, similar to jev

github.com2026-09-17authorityHN 7 pts · 1 comments2,103 words

An open-source implementation of Jev option scoring runs on local Gemma 3 4B models using MLX on Apple silicon or PyTorch on other platforms. The model achieves efficient option scoring by prefilling the context once and expanding the KV cache across a batch of options in a single forward pass without decoding. The score is the log-probability of the option tokens given the context; a softmax over the option scores gives a probability per option, like jevlike-predict.written, reviewed, cited

Frozen-feature extraction and head training workflows currently require MLX on Apple silicon.
PMI normalization requires an extra batched pass to subtract the option's unconditional log-prob.

Show HN: Gemma 3 inference in pure C++ with Metal acceleration

github.com2026-07-04authorityabout the gemma 3 familyHN 6 pts · 2 comments193 words

MetalChat is a Metal-accelerated C++ framework and command line interpreter for inference of Meta Llama, Google Gemma, and Liquid models. The library API and CLI are under active development, therefore they may change without any deprecation notice. The framework is distributed under GPLv3 license and can be installed using Homebrew package manager with brew tap ybubnov/metalchat.written, reviewed, cited

Fine-tune Google's Gemma 3

unsloth.ai2025-03-19authorityHN front pageabout the gemma 3 familyHN 226 pts · 78 comments

Gemma 3 Technical Report [pdf]

storage.googleapis.com2025-03-12HN front pageabout the gemma 3 familyHN 488 pts · 245 comments

I have written gemma3 inference in pure C

github.com2026-01-26authorityabout the gemma 3 familyHN 65 pts · 23 comments2,207 words

Gemma3 – The current strongest model that fits on a single GPU

ollama.com2025-03-12HN front pageabout the gemma 3 familyHN 252 pts · 138 comments

All coverage tracked for this model (26) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
Open-jev: One-pass option scoring with Gemma 3 4B, similar to jev	github.com	2026-09-17	2026-10-07	hn
Show HN: Gemma 3 inference in pure C++ with Metal acceleration	github.com	2026-07-04	2026-10-07	hn
Selectively reducing eval awareness and murder in Gemma 3 27B via steering	lesswrong.com	2026-03-12	2026-10-07	hn
I have written gemma3 inference in pure C	github.com	2026-01-26	2026-10-07	hn
Own your AI: Learn how to fine-tune Gemma 3 270M and run it on-device	developers.googleblog.com official	2025-10-10	2026-10-07	hn
Gemma 3 270M re-implemented in pure PyTorch for local tinkering	github.com	2025-08-20	2026-10-07	hn
Landrecords – cheap nationwide parcel dataset standardized using gemma3	landrecords.us	2025-08-20	2026-10-07	hn
Gemma 3 270M: Compact model for hyper-efficient AI	developers.googleblog.com official	2025-08-14	2026-10-07	hn
I Gave Up Trying to Run Gemma 3 27B on AWS P3dn.24xlarge	medium.com	2025-08-11	2026-10-07	hn
Gemma 3n: The Developer Guide	simonwillison.net	2025-06-27	2026-10-07	hn
Gemma 3n available in the open-source ecosystem	huggingface.co	2025-06-26	2026-10-07	hn
Introducing Gemma 3n	developers.googleblog.com official	2025-06-26	2026-10-07	hn
A C++ library to efficiently run Gemma-3N across various platform	github.com	2025-06-11	2026-10-07	hn
Gemma 3n Architectural Innovations – Speculation and poking around in the model	old.reddit.com	2025-05-25	2026-10-07	hn
Gemma 3n preview: Mobile-first AI	developers.googleblog.com official	2025-05-20	2026-10-07	hn
Gemma 3 QAT Models	simonwillison.net	2025-04-19	2026-10-07	hn
QAT optimized Gemma 3 slash VRAM needs 54GB –> 14.1GB while maintaining quality	old.reddit.com	2025-04-19	2026-10-07	hn
Gemma 3 QAT Models: Bringing AI to Consumer GPUs	developers.googleblog.com official	2025-04-18	2026-10-07	hn
Gemma3 Function Calling	ai.google.dev official	2025-03-23	2026-10-07	hn
Google calls Gemma 3 the most powerful AI model you can run on one GPU	theverge.com	2025-03-20	2026-10-07	hn
Fine-tune Google's Gemma 3	unsloth.ai	2025-03-19	2026-10-07	hn
Google's Gemma 3 is an open source, multimodal single-GPU AI 128K context window	arstechnica.com	2025-03-13	2026-10-07	hn
Gemma3 – The current strongest model that fits on a single GPU	ollama.com	2025-03-12	2026-10-07	hn
Gemma 3: The most capable model you can run on a single GPU or TPU	blog.google official	2025-03-12	2026-10-07	hn
Gemma 3: Google's new multimodal models	ai.google.dev official	2025-03-12	2026-10-07	hn
Gemma 3 Technical Report [pdf]	storage.googleapis.com	2025-03-12	2026-10-07	hn

Papers

8 papers

Papers 8

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).

PIQA: Reasoning about Physical Commonsense in Natural Language cited in card peer-reviewed · measured

Yonatan Bisk, Rowan Zellers, Ronan Le Bras +2 · 2019-11-26 · Proceedings of the AAAI Conference on Artificial Intelligence · PDF · HF paper page (5 upvotes) · evidence
TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension cited in card peer-reviewed · measured

Mandar Joshi, Eunsol Choi, Daniel S. Weld +1 · 2017-05-09 · Annual Meeting of the Association for Computational Linguistics (ACL) · PDF · HF paper page (2 upvotes) · evidence
On the Cross-lingual Transferability of Monolingual Representations cited in card peer-reviewed · measured

Mikel Artetxe, Sebastian Ruder, Dani Yogatama · 2019-10-25 · Annual Meeting of the Association for Computational Linguistics (ACL) · PDF · HF paper page · evidence
Gemini: A Family of Highly Capable Multimodal Models cited in card

Gemini Team, Rohan Anil, Sebastian Borgeaud +1348 · 2023-12-19 · PDF · HF paper page (50 upvotes)
MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI cited in card

Xiang Yue, Yuansheng Ni, Kai Zhang +19 · 2023-11-27 · PDF · HF paper page (40 upvotes)
GPQA: A Graduate-Level Google-Proof Q&A Benchmark cited in card

David Rein, Betty Li Hou, Asa Cooper Stickland +5 · 2023-11-20 · PDF · HF paper page (38 upvotes)
BLINK: Multimodal Large Language Models Can See but Not Perceive cited in card

Xingyu Fu, Yushi Hu, Bangzheng Li +7 · 2024-04-18 · PDF · HF paper page (27 upvotes)
Evaluating Large Language Models Trained on Code cited in card

Mark Chen, Jerry Tworek, Heewoo Jun +55 · 2021-07-07 · PDF · HF paper page (12 upvotes)

Similar models

3 models

Similar models

Qwen3-VL-4B-Instruct · 4.44B
same size band (M), same task (image-text-to-text), more providers, permissive license
medgemma-4b-it · 4.30B
same size band (M), same task (image-text-to-text)
Qwen3.5-4B · 4.66B
same size band (M), same task (image-text-to-text), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:38 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:09 UTC · Last verified 2026-10-25

Follow gemma-3-4b-pt

Watch this model

/feeds/models/google--gemma-3-4b-pt.xml(RSS 2.0) · every change to gemma-3-4b-pt
/feeds/families/google--gemma-3.xml(RSS 2.0) · every change across the gemma-3 family

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