translategemma-4b-it

google Gemma Terms of Use out 9 months ago

4.30B parameters, images and text in, text out, 128K tokens of context.

Verified 2026-10-25 Rank #333 in LLMs · score 81.8 method v1 Sources 3 Downloads 30d 13.4k · #485 of 619 tracked

Changed this week: 1 new quantized build on the Hub (45 in all).2026-10-24

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?13.4kdownloads, 30 days#485 of 619 we track · below the median+10 likes, downloads −46% this week886 likesHugging Face2026-10-24

Worth reading

Picked by rule from 7 tracked pieces: the lab's own post, the most shared, a field report, the latest.

TranslateGemma inherits multimodal processor expectations from the PaliGemma lineage, requiring the pillow library even in a text-focused workflow.

From Fine-Tuning TranslateGemma-4B for Better Welsh Translations on an H200 GPU · metalglot.com · 2026-05-13

Latest · 5 months ago

Fine-Tuning TranslateGemma-4B for Better Welsh Translations on an H200 GPU

A guide details fine-tuning TranslateGemma-4B for Welsh on a single NVIDIA H200 GPU to improve translation while preserving instruction-following capability.

TranslateGemma inherits multimodal processor expectations from the PaliGemma lineage, requiring the pillow library even in a text-focused workflow.
The training loop uses completion-only loss to keep optimization pressure on the model's answer instead of penalizing it for prompt tokens.

metalglot.com 2026-05-13 1 HN pts · 1 comments 4.9k words

7 best of 7 tracked

The 1 pick above, plus these 6. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleTranslateGemma: A new suite of open translation modelsblog.google · 2026-01-15 · 11 HN pts · official
ArticleComparing the best open source TranslateGemma projectsmetalglot.com · 2026-05-03 · 2 HN pts
ArticleTranslateGemma 4B：评测、参数、下载与模型卡 | DataLearnerAIdatalearner.com · undated
Articletranslategemma-4b-it huggingface.co api & google translategemma-4b-it github AI Model - Toolifytoolify.ai · undated
ArticleTranslateGemma: Proven Setup, 9 Benchmarks, Local Deploybinaryverseai.com · 2026-01-20
ArticleTranslateGemma Complete Guide: Benchmarks, Setup & Usage (2026) | AiCybr Blogaicybr.com · 2026-01-19

Show all 6

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: 45 quantized builds and 20 Spaces.

As a service0no API provider listed
On own hardware45quantized builds
Built on4126 fine-tunes · 14 adapters · 1 merges
In apps20Hugging Face Spaces

Top Spaces running it: fantos/translate-gemma ♥ 78 · webml-community/TranslateGemma-WebGPU ♥ 53 · hysts/translategemma-27b-it ♥ 34 · gnumanth/translategemma-4b-it ♥ 5

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.

Most coverage came in the first two weeks

Teal is the weights (2026-01-12), amber triangles are 5 dated articles (each links to it), grey ticks are repo commits.

What people built on it

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

86 derivatives on the Hub, most of them quantized

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

GGUF mradermacher/translategemma-4b-it-GGUF for llama.cpp, Ollama, LM Studio 281k downloads MLX mlx-community/translategemma-4b-it-4bit for Apple silicon 32.1k downloads NVFP4 kaitchup/translategemma-4b-it-NVFP4 for vLLM, SGLang 81.8k downloads

All variants →

Reference

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

Benchmarks

2 rows

Benchmarks 2 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (2)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 Vistra* (4 langs) MetricX ↓	2.57	4B	model author	2026-10-07
author-reported WMT24++ · Work across languages	5.32	4B	model author	2026-10-07

What the card says

2 cited sentences

Summary generated, cited

TranslateGemma is a 4.97B parameter image-text-to-text model based on the Gemma 3 family.^[1]^[2] The model performs translation tasks across 55 languages using text or image inputs.^[3]

parameters human
“based on the Gemma 3 family”
“designed to handle translation tasks across 55 languages”

Generated 2026-10-07 15:08 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

5 use cases

What it is good at 5 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

translategemma-4b-it has 4.30B parameters. (safetensors metadata, observed 2026-10-24)
Context length is 131,072 tokens (128K). (config.json max_position_embeddings, 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 2026-01-12, last modified 2026-01-28. (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
Context length	131,072 (config.json max_position_embeddings)
Gated	manual
Papers	arXiv:2601.09012, arXiv:2503.19786
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 translategemma-4b-it: 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/translategemma 86 derivatives on the Hub

quantized: 45merge: 1finetune: 26adapter: 14

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
mradermacher/translategemma-4b-it-GGUF	quantized	GGUF	—	—	gemma	mradermacher	67	281k
kaitchup/translategemma-4b-it-NVFP4	quantized	NVFP4	3.57B	M	gemma	kaitchup	0	81.8k
bullerwins/translategemma-4b-it-GGUF	quantized	GGUF	—	—	gemma	bullerwins	5	60.8k
SandLogicTechnologies/translategemma-4b-it-GGUF	quantized	GGUF	—	—	gemma	SandLogicTechnologies	2	41.7k
mlx-community/translategemma-4b-it-4bit	quantized	-4bit	3.88B	M	gemma	mlx-community	11	32.1k
mradermacher/translategemma-4b-it-i1-GGUF	quantized	GGUF	—	—	gemma	mradermacher	4	19.2k
NikolayKozloff/translategemma-4b-it-Q8_0-GGUF	quantized	Q8_0	—	—	gemma	NikolayKozloff	12	13.1k
kaitchup/translategemma-4b-it-FP8-Dynamic	quantized	-FP8	4.30B	M	gemma	kaitchup	1	11.4k
tatsuyaaaaaaa/translategemma-4b-it-gguf	quantized	gguf	—	—	—	tatsuyaaaaaaa	0	10.2k
mlx-community/translategemma-4b-it-4bit_immersive-translate	quantized	-4bit	3.88B	M	gemma	mlx-community	11	7.6k
mlx-community/translategemma-4b-it-8bit	quantized	-8bit	3.88B	M	gemma	mlx-community	5	5.8k
prithivMLmods/translategemma-4b-it-f32-GGUF	quantized	GGUF	—	—	gemma	prithivMLmods	1	5.7k
Qwe1325/translategemma-4b-it-GGUF	quantized	GGUF	—	—	—	Qwe1325	2	5.3k
DevQuasar/google.translategemma-4b-it-GGUF	quantized	GGUF	—	—	—	DevQuasar	2	5.1k
42ailab/TranslateGemma-4B-GGUF	quantized	GGUF	—	—	gemma	42ailab	4	4.4k
Gunulhona/Gemma-3-4B	merge	—	4.30B	M	—	Gunulhona	1	4k
bingbangboom/translategemma-4b-it-Q4_K_M-GGUF	quantized	GGUF	—	—	gemma	bingbangboom	0	3.2k
bRadu/translategemma-4b-it-novision	finetune	—	3.88B	M	other	bRadu	3	1.9k
aoiandroid/translategemma-4b-it-GGUF	quantized	GGUF	—	—	gemma	aoiandroid	0	2k
vjchou/translategemma-4b-it-GGUF	quantized	GGUF	—	—	gemma	vjchou	0	1.9k
matrixportalx/translategemma-4b-it-Q5_K_M-GGUF	quantized	GGUF	—	—	gemma	matrixportalx	0	1.5k
azmnc/translategemma-4b-it-Q4_K_M_GGUF	quantized	GGUF	—	—	gemma	azmnc	0	1.1k
litert-community/TranslateGemma-4B-IT	finetune	—	—	—	gemma	litert-community	21	0
light434/translategemma-4b-it-int8-ov	quantized	int8	—	—	gemma	light434	2	937
smdesai/translategemma-4b-it-4bit	quantized	-4bit	3.88B	M	gemma	smdesai	0	956
DeltaKX/translategemma-4b-it-Q4_K_M-GGUF	quantized	GGUF	—	—	gemma	DeltaKX	0	861
Yacinedh/translategemma-4b-khutbah	quantized	GGUF	4.30B	M	gemma	Yacinedh	0	826
ArmanEspiar/TranslateGemma-4b-PersiaFT-v1	quantized	GGUF	4.97B	M	apache-2.0	ArmanEspiar	0	822
zanish-labs/translategemma-4b-it-Q2_K-gguf	quantized	Q2_K	—	—	gemma	zanish-labs	2	690
aidsoid/translategemma-4b-it-4bit_immersive-translate	quantized	-4bit	3.88B	M	gemma	aidsoid	2	648
alhasanmuhammad/translategemma-4b-it-21hrs	finetune	—	4.30B	M	gemma	alhasanmuhammad	0	729
Skyline23/translategemma-4b-it-coreml	quantized	—	—	—	gemma	Skyline23	1	627
AdrienGuille/GemmAMR-fr-v1	finetune	—	4.30B	M	—	AdrienGuille	1	506
barakplasma/translategemma-4b-it-android-task-quantized	finetune	-quant	—	—	other	barakplasma	9	61
lyf07/Translategemma-4B-it-WALAR	finetune	—	769K	S	mit	lyf07	4	251
matrixportalx/translategemma-4b-it-Q4_K_M-GGUF	quantized	GGUF	—	—	gemma	matrixportalx	0	422
helvetia/translategemma-4b-rm-sutsilv-lora-v1	adapter	—	—	—	—	helvetia	0	422
bingbangboom/translategemma-4b-it-Q6_K-GGUF	quantized	Q6_K	—	—	gemma	bingbangboom	0	414
OpenOranje/translategemma-4b-it-mlx-4bit	quantized	mlx	3.88B	M	gemma	OpenOranje	0	391
fotiecodes/Turaco-gem-mt-en-wes	adapter	GGUF	4.30B	M	gemma	fotiecodes	0	375

Repository changelog

2 commits

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

Repository changelog 2 commits

Date	Commit	SHA
2026-01-28	Update generation_config.json	10042cb
2026-01-15	Super-squash branch 'main' using huggingface_hub	4eded5e

Articles and coverage

7 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.

Official1

TranslateGemma: A new suite of open translation models

blog.google2026-01-15officialauthorityabout the translategemma familyHN 11 pts · 1 comments644 words

Google introduces TranslateGemma, a suite of open translation models built on Gemma 3 available in 4B, 12B, and 27B parameter sizes. The 12B model outperforms the Gemma 3 27B baseline as measured using MetricX on the WMT24++ benchmark through a specialized two-stage fine-tuning process. Deployment options include mobile edge, consumer laptops, or a single H100 GPU or TPU in the cloud.written, reviewed, cited

The team pushed the boundaries by training on nearly 500 additional language pairs beyond the core 55 language pairs.

Analysis6

Fine-Tuning TranslateGemma-4B for Better Welsh Translations on an H200 GPU

metalglot.com2026-05-13HN 1 pts · 1 comments4,916 words

A guide details fine-tuning TranslateGemma-4B for Welsh on a single NVIDIA H200 GPU to improve translation while preserving instruction-following capability. A 5% pilot run completes in roughly 40 minutes using a 70:30 translation-to-instruction mix, with a full 1.35M-row dataset taking approximately 33 hours on one GPU. The setup requires CUDA 12.8 and leverages Flash Attention 3 to maintain a 2048-token context window.written, reviewed, cited

TranslateGemma inherits multimodal processor expectations from the PaliGemma lineage, requiring the pillow library even in a text-focused workflow.
The training loop uses completion-only loss to keep optimization pressure on the model's answer instead of penalizing it for prompt tokens.

Comparing the best open source TranslateGemma projects

metalglot.com2026-05-03about the translategemma familyHN 2 pts · 0 comments2,487 words

This guide categorizes open-source TranslateGemma tools by specific use cases rather than general hype. fastapi-gemma-translate offers a robust REST API backend under the MIT License, while Llingua provides a browser-only approach using WebAssembly. AI Subtitle Tools combine FFmpeg audio extraction for media pipelines, though most tools struggle with structured localization formats like i18next JSON or XLIFF.written, reviewed, cited

TOST combines OCR and local inference to translate screenshots and desktop software on-screen.
TranslateGemma Studio handles long-form text and chunking better than other local web frontends.
petit_trad targets technically confident users who want fine control over execution backends including CUDA, Metal, Vulkan, and CPU-only modes.

TranslateGemma 4B：评测、参数、下载与模型卡 | DataLearnerAI

datalearner.com187 words

Google DeepMind released TranslateGemma 4B in January 2026 as the smallest model in its TranslateGemma family built on Gemma 3. The 4B parameter model translates across 55 languages and supports image-to-text translation while optimizing performance for mobile and edge devices.generated, cited

translategemma-4b-it huggingface.co api & google translategemma-4b-it github AI Model - Toolify

toolify.ai

TranslateGemma: Proven Setup, 9 Benchmarks, Local Deploy

binaryverseai.com2026-01-20about the translategemma family2,492 words

TranslateGemma Complete Guide: Benchmarks, Setup & Usage (2026) | AiCybr Blog

aicybr.com2026-01-19about the translategemma family1,590 words

Papers

2 papers

Papers 2

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).

TranslateGemma Technical Report this model

Mara Finkelstein, Isaac Caswell, Tobias Domhan +18 · 2026-01-13 · PDF · HF paper page (23 upvotes)
Gemma 3 Technical Report cited in card

Gemma Team, Aishwarya Kamath, Johan Ferret +213 · 2025-03-25 · PDF · HF paper page (59 upvotes)

Similar models

3 models

Similar models

gemma-3-4b-it · 4.30B
same size band (M), same task (image-text-to-text), more providers
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)

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:37 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:08 UTC · Last verified 2026-10-25

Follow translategemma-4b-it

Watch this model

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

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