gemma-3-4b-it

Home/Models/google/gemma-3-4b-it

gemma-3-4b-it

google/gemma-3-4b-it

kind: original
quant: bf16 (16-bit)
family: gemma
modality: image-text-to-text
root model: google/gemma-3-4b-pt
provider: google
license: gemma
language: —
context: —
library: transformers
vRAM (GB): —
size on disk: —
downloads: 2,093,730
likes: 1,432
created: —
last modified: 2025-03-21T20:20:53+00:00 (2025)
huggingface: https://huggingface.co/google/gemma-3-4b-it
tags: transformerssafetensorsgemma3image-text-to-textconversationalarxiv:1905.07830arxiv:1905.10044arxiv:1911.11641arxiv:1904.09728arxiv:1705.03551arxiv:1911.01547arxiv:1907.10641arxiv:1903.00161arxiv:2009.03300arxiv:2304.06364arxiv:2103.03874arxiv:2110.14168arxiv:2311.12022arxiv:2108.07732arxiv:2107.03374arxiv:2210.03057arxiv:2106.03193arxiv:1910.11856arxiv:2502.12404

Benchmark scores (20)

benchmark	category	score	context	conf	evidence
PIQA	knowledge	73.8	[PIQA][]	medium	`\| [PIQA][piqa] \| 0-shot \| 73.8 \| 79.6 \| 81.8 \| 83.3 \|`
DocVQA	multimodal	72.8	[DocVQA][] (val)	medium	`\| [DocVQA][docvqa] (val) \| 72.8 \| 82.3 \| 85.6 \|`
ChartQA	multimodal	63.6	[ChartQA][]	medium	`\| [ChartQA][chartqa] \| 63.6 \| 74.7 \| 76.3 \|`
BoolQ	knowledge	63.2	[BoolQ][]	medium	`\| [BoolQ][boolq] \| 0-shot \| 63.2 \| 72.3 \| 78.8 \| 82.4 \|`
HellaSwag	knowledge	62.3	[HellaSwag][]	medium	`\| [HellaSwag][hellaswag] \| 10-shot \| 62.3 \| 77.2 \| 84.2 \| 85.6 \|`
MMLU	knowledge	59.6	[MMLU][]	medium	`\| [MMLU][mmlu] \| 5-shot \| 59.6 \| 74.5 \| 78.6 \|`
WinoGrande	knowledge	58.2	[WinoGrande][]	medium	`\| [WinoGrande][winogrande] \| 5-shot \| 58.2 \| 64.7 \| 74.3 \| 78.8 \|`
MBPP	coding	46.0	[MBPP][]	medium	`\| [MBPP][mbpp] \| 3-shot \| 46.0 \| 60.4 \| 65.6 \|`
DROP	reasoning	42.4	[DROP][]	medium	`\| [DROP][drop] \| 1-shot \| 42.4 \| 60.1 \| 72.2 \| 77.2 \|`
AGIEval	reasoning	42.1	[AGIEval][]	medium	`\| [AGIEval][agieval] \| 3-5-shot \| 42.1 \| 57.4 \| 66.2 \|`
TriviaQA	knowledge	39.8	[TriviaQA][]	medium	`\| [TriviaQA][triviaqa] \| 5-shot \| 39.8 \| 65.8 \| 78.2 \| 85.5 \|`
MMMU	multimodal	39.2	[MMMU][] (pt)	medium	`\| [MMMU][mmmu] (pt) \| 39.2 \| 50.3 \| 56.1 \|`
GSM8K	math	38.4	[GSM8K][]	medium	`\| [GSM8K][gsm8k] \| 8-shot \| 38.4 \| 71.0 \| 82.6 \|`
HumanEval	coding	36.0	[HumanEval][]	medium	`\| [HumanEval][humaneval] \| 0-shot \| 36.0 \| 45.7 \| 48.8 \|`
MMLU	knowledge	29.2	[MMLU][] (Pro COT)	medium	`\| [MMLU][mmlu] (Pro COT) \| 5-shot \| 29.2 \| 45.3 \| 52.2 \|`
BBH	reasoning	28.4	[BIG-Bench Hard][]	medium	`\| [BIG-Bench Hard][bbh] \| few-shot \| 28.4 \| 50.9 \| 72.6 \| 77.7 \|`
MMLU	knowledge	24.9	[Global-MMLU-Lite][global--lite]	medium	`\| [Global-MMLU-Lite][global-mmlu-lite] \| 24.9 \| 57.0 \| 69.4 \| 75.7 \|`
MATH	math	24.2	[MATH][]	medium	`\| [MATH][math] \| 4-shot \| 24.2 \| 43.3 \| 50.0 \|`
GPQA	reasoning	15.0	[GPQA][]	medium	`\| [GPQA][gpqa] \| 5-shot \| 15.0 \| 25.4 \| 24.3 \|`
MGSM	math	2.04	[MGSM][]	medium	`\| [MGSM][mgsm] \| 2.04 \| 34.7 \| 64.3 \| 74.3 \|`

Discussions (90)

title	author	status	type
Create .eval_results/ask-my-agent-bench-2.yaml	sapbot	open	disc
Create .eval-results/ask-my-agent-bench-2.yaml	sapbot	closed	disc
VORTEXRAG: 7-Layer RAG — EM 74.8 on QA benchmarks, solves Semantic Drift [open source]	vigneshwar234	closed	disc
Add EvalEval community eval results	EvalEvalBot	open	disc
Update README.md	Mohamed0mer	open	disc
I built a unified wrapper for llmcompressor, llama.cpp & coremltools. Looking for LLM users to help me break it!	kinderasteroid	open	disc
Finetuning Code Link In Native PyTorch	dutta18	open	disc
anis	Aniskhan29	draft	disc
anis	Aniskhan29	draft	disc
Request: DOI	AlperIsleyen	closed	disc
Request: DOI	Salehsst	open	disc
Add Artificial Analysis evaluations for gemma-3-4b	davidlms	open	disc
Add model-index with comprehensive benchmark evaluations	davidlms	open	disc
Test	Kishi1	draft	disc
Token Count Calculation in SFT Data Distribution Curation	tcy006	open	disc
gemma 3	bojupexa	draft	disc
LoRA/PEFT can’t target nn.Parameter vision→text projection in Gemma-3 VLM — why use nn.Parameter instead of nn.Linear(bias=False)?	alexanderyj	open	disc
4bit Quantization Failure for gemma 4b	shabha7092	open	disc
Gemma full	RamarajuK	draft	disc
test	munbujang	draft	disc
Inference Issue with Multi GPU's	m-nameer	open	disc
<pad> <pad> <pad> fix	test53100	open	disc
processor does not have a chat template	Prabhjot410	open	disc
onnx model	Gerald001	open	disc
Parameters are contradictory use_cache=False	lemon0703	open	disc
onnx model	robustdev	closed	disc
confusion with model import	Charushila	closed	disc
Re: failing structured output generation	msi-sbraun-11	open	disc
Gemma3-U1	ForSureTesterSim	closed	disc
Cannot recreate benchmark results	alternis	open	disc
The example code provided (without modifications) produces gibberish	zqQH	open	disc
Update README.md	Feelthevenom	closed	disc
Solve error about "dynamo"	audrey666	closed	disc
Update README.md	dj1507	open	disc
Missing values in 4B, 12B config.json	depasquale	open	disc
git-clone	devrf	draft	disc
Update preprocessor_config.json	nm-research	open	disc
use model run failed	pb-1	open	disc
cache_dir with ssd super slow and once loaded in nvidia quadro P4000 with 4bit return nothing Image is made of multiple images inside with timestamp	jef06	closed	disc
Bug report: torch._dynamo.exc.Unsupported: Unexpected type in sourceless builder transformers.models.gemma3.configuration_gemma3.Gemma3TextConfig	YovelRVeritex	open	disc
Update config.json	olegshulyakov	open	disc
Update README.md	VANNVISAL	open	disc
Tokenizer's `model_max_length` unusual value	awni	closed	disc
Update README.md	DimaYastrebov	closed	disc
Error in loading Model	sans07	open	disc
gemma3-1b-it	TAENNY	draft	disc
Explicitly include num_attention_heads and num_key_value_heads in text_config for Gemma3 4B	ebsmothers	open	disc
Deployment issues on sagemaker	tychen6677	open	disc
Please open-source Gemini 1.5 Flash	drguolai	open	disc
CUDA error: device-side assert triggered	ArshiaSoori	open	disc
Hih	Nanoser12	closed	disc
Low GPU Utilization during inference?	BagelBig	open	disc
Object detection capabilities ?	syrineM	open	disc
SigLIP or SigLIP2 encoder?	orrzohar	open	disc
text-generation	rakmik	open	disc
I'm releasing the speech version of Gemma-3!	junnei	open	disc
Update README.md	sachin7yadava	open	disc
Issue with vLLM Deployment of gemma-3-4b-it on Tesla T4 - No Output	twodaix	open	disc
Batch processing on a GPU?	buckeye17-bah	open	disc
Please release also Gemini Flash 1.5 weights.	ZeroWw	open	disc
Finetuning script using HuggingFace	2U1	closed	disc
VRAM not freed during long generations (Gemma, max_new_tokens=3000)	Nessit	open	disc
The checkpoint you are trying to load has model type `gemma3` but Transformers does not recognize this architecture	bajajrahul03	open	disc
Does not works with bitsandbytes 4bit and 8bit	zokica	open	disc
Pan&Scan is desable by default ?	Jack-HR-Red	open	disc
Update README	patloeber	merged	disc
Update README.md	merve	merged	disc
Upload tokenizer.model	Nexesenex	open	disc
cannot import name 'Gemma3ForConditionalGeneration' from 'transformers'	ansh10dave	closed	disc
Having a hard time fine tuning with LoRA	abcampbell	open	disc
Adding "vocab_size" to config.json	nmysore	open	disc
The checkpoint you are trying to load has model type `gemma3` but Transformers does not recognize this architecture.	bedoonraj	open	disc
getting key error while running inference for pipeline api given in "Usage section"	himanshu-bluebrain	open	disc
How can I run the pipeline without the Image input?	prakhaaaas	open	disc
Model Produces no Outputs	RonanMcGovern	closed	disc
AttributeError: 'HybridCache' object has no attribute 'float'	naruto-soop	open	disc
Not enough parameters in 4b config.json	krammnic	open	disc
AttributeError: 'Gemma3Config' object has no attribute 'vocab_size'	MinecAnton209	open	disc
Add video inference	merve	open	disc
Failed to deploy on Amazon SageMaker ml.g5.4xlarge instance	Khalizo	closed	disc
Local Installation Video and Testing - Step by Step	fahdmirzac	open	disc
Question About Benchmark Version in README	lianghsun	open	disc
Feedback: mildly disappointed there isn't 9B version	nonetrix	open	disc
I am getting error for shape mismatch with the provided code	abhgarg	closed	disc
Update chat_template	Xenova	merged	disc
Update README.md	merve	merged	disc
Add code snippet, change pipeline tag and library name	merve	closed	disc
Add code snippet, change pipeline tag and library name	merve	closed	disc
Add <end_of_turn> as a terminator	pcuenq	closed	disc
Upload tokenizer	Xenova	merged	disc