Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › t5gemma-2 family

t5gemma-2-270m-270m

Hugging Face ↗ Follow changes

google Gemma Terms of Use out 12 months ago

786M parameters, images and text in, text out.

Verified 2026-10-25 Rank #586 in LLMs · score 74.3 method v1 Sources 3 Downloads 30d 59.5k · #367 of 619 tracked

Can I run it?0.5GB 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?59.5kdownloads, 30 days#367 of 619 we track · below the median210 likesHugging Face2026-10-24

Worth reading

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

Latest · 10 months ago

T5Gemma 2: The next generation of encoder-decoder models

T5Gemma 2 launches as the first multi-modal and long-context encoder-decoder models based on Gemma 3, featuring tied embeddings and merged attention mechanisms.

  • The release explicitly states that no post-trained or instruction-tuned checkpoints will be made available to the public.
  • Developers must perform their own supervised fine-tuning because the provided pre-trained checkpoints are designed specifically for post-training.
blog.google 2025-12-18 153 HN pts · 28 comments 619 words about the t5gemma 2 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: 3 Spaces.

  • As a service0no API provider listed
  • On own hardware0quantized builds
  • Built on1815 fine-tunes · 3 adapters · 0 merges
  • In apps3Hugging Face Spaces

Top Spaces running it: KwabsHug/GameConfigIdea ♥ 17

Will it fit?

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

Fits a 16 GB card at INT4

About 0.5 GB at INT4, 0.9 GB at FP8, 1.9 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 0.5 GB of weights (estimate; KV cache and activations extra)INT40.5 GBFP8: about 0.9 GB of weights (estimate; KV cache and activations extra)FP80.9 GBBF16: about 1.9 GB of weights (estimate; KV cache and activations extra)BF161.9 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 0.5 GB of weights (estimate; KV cache and activations extra)INT40.5 GBFP8: about 0.9 GB of weights (estimate; KV cache and activations extra)FP80.9 GBBF16: about 1.9 GB of weights (estimate; KV cache and activations extra)BF161.9 GB

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 13 repo commits so far

Teal is the weights (2025-10-25), amber triangles are 1 dated article (each links to it), grey ticks are repo commits.

NovDecJanFebMarAprMayJunJulAugSepOct2025-12-16: 1 repo commit2025-12-15: 1 repo commit2025-12-12: 5 repo commits2025-12-11: 1 repo commit2025-11-25: 2 repo commits2025-10-25: 3 repo commitsT5Gemma 2: The next generation of encoder-decoder models · 2025-12-18Weights 2025-10-25 · 2025-10-25Weights 2025-10-25now
NovDecJanFebMarAprMayJunJulAugSepOct2025-12-16: 1 repo commit2025-12-15: 1 repo commit2025-12-12: 5 repo commits2025-12-11: 1 repo commit2025-11-25: 2 repo commits2025-10-25: 3 repo commitsT5Gemma 2: The next generation of encoder-decoder models · 2025-12-18Weights 2025-10-25 · 2025-10-25Weights 2025-10-25now

What people built on it

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

18 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.

15 fine-tunes (Hugging Face model tree)3 adapters (Hugging Face model tree)15 fine-tunes3 adapters
15 fine-tunes (Hugging Face model tree)3 adapters (Hugging Face model tree)15 fine-tunes3 adapters

All variants →

Reference

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

Benchmarks

23 rows

Benchmarks 23 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (23)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.
BenchmarkValueSetting / notesReported byDate
author-reported AGIEval20.93-5-shotmodel author2026-10-07
author-reported ARC Challenge · Chat and follow instructions27.925-shotmodel author2026-10-07
author-reported BBH · Reason through hard problems22.9few-shotmodel author2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions57.40-shotmodel author2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions66.50-shotmodel author2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions46.10-shotmodel author2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions14.85-shotmodel author2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions35-shotmodel author2026-10-07
author-reported Commonsense QA sets · Chat and follow instructions57.90-shotmodel author2026-10-07
author-reported DROP39.41-shotmodel author2026-10-07
author-reported FLORES · Work across languages23.9—model author2026-10-07
author-reported GPQA Diamond · Reason through hard problems9.65-shotmodel author2026-10-07
author-reported GSM8K · Reason through hard problems1.78-shotmodel author2026-10-07
author-reported Global-MMLU · Work across languages23.5—model author2026-10-07
author-reported HellaSwag · Chat and follow instructions41.110-shotmodel author2026-10-07
author-reported HumanEval · Write code with agents4.30-shotmodel author2026-10-07
author-reported MATH-500 · Reason through hard problems1.54-shotmodel author2026-10-07
author-reported MBPP · Write code with agents5.63-shotmodel author2026-10-07
author-reported MGSM · Work across languages1.8—model author2026-10-07
author-reported MMLU-Pro · Reason through hard problems10.65-shotmodel author2026-10-07
author-reported WMT24++ · Work across languages26.9—model author2026-10-07
author-reported Winogrande · Chat and follow instructions53.95-shotmodel author2026-10-07
author-reported XQuAD (all)40.8—model author2026-10-07

What the card says

1 cited sentence

Summary generated, cited

The architecture consists of a 270M encoder and 270M decoder.[1]

  1. “270M encoder and 270M decoder”

Generated 2026-10-07 15:13 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

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

Pipeline tagimage-text-to-text
Librarytransformers
ArchitecturesT5Gemma2ForConditionalGeneration
Model typet5gemma2
Dtypebfloat16
Params by dtypeBF16: 786,029,296
Size bandS
MoEno
Gatedmanual
PapersarXiv: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:2502.12404, arXiv:2106.03193, arXiv:1910.11856, arXiv:2104.12756, arXiv:2311.16502, arXiv:2203.10244, arXiv:1810.12440, arXiv:1908.02660, arXiv:2404.06654, arXiv:2409.12640
LicenseGemma Terms of Use (gemma)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: not in registry · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: not in registry · 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 t5gemma-2-270m-270m: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)1.9 GB · Edge (<4 GB)
FP8 (1 byte/param)0.9 GB · Edge (<4 GB)
INT4 (0.5 bytes/param)0.5 GB · Edge (<4 GB)

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

All variants

18 variants

Lineage

Original model Family: google/t5gemma-2 18 derivatives on the Hub

finetune: 15adapter: 3

Top variants (18 shown, ranked by downloads)

Relation
18 rows
Variant Relation Quant Params Band License Author Likes Downloads
KaLM-Embedding/KaLM-Reranker-V1-Nano finetune — 786M S apache-2.0 KaLM-Embedding 10 2.9k
tomaarsen/t5gemma2-270m-gooaq-cmnrl finetune — 686M S apache-2.0 tomaarsen 1 722
knowledgator/t5gemma-2-text-encoder-270m finetune — 268M S gemma knowledgator 1 696
KaLM-Embedding/KaLM-Reranker-V1-Nano-R2 finetune — 786M S apache-2.0 KaLM-Embedding 2 534
jadanovitch/t5gemma-2-270m-encoder finetune — 268M S gemma jadanovitch 0 366
PhatcatDK/t5gemma-2-270m-270m-encoder-only finetune — 686M S gemma PhatcatDK 0 347
kingabzpro/t5gemma2-latex-ocr-1k finetune — 786M S gemma kingabzpro 2 159
BharathP08/t5gemma-270m-legal-v1 adapter — — — gemma BharathP08 1 121
mindchain/t5gemma-270m-container-status finetune — 786M S — mindchain 1 102
evatsirmi/t5gemma_greek_lora adapter — — — gemma evatsirmi 0 124
Pieces/time-mapping-t5gemma-270m-best finetune — 786M S apache-2.0 Pieces 1 74
mindchain/t5gemma2-sae-all-layers finetune — — — mit mindchain 2 0
la0ka1/T5Gemma-2-270M-OWTdistilled finetune — — — gemma la0ka1 2 0
uthayamurthy/origin-task1-t5gemma2-270m finetune — 786M S gemma uthayamurthy 0 86
Pieces/time-mapping-t5gemma-270m-lr5e5 finetune — 786M S apache-2.0 Pieces 0 74
kingabzpro/t5gemma-2-latex-ocr finetune — — — — kingabzpro 1 0
nakasyou/saikyo-ime-reranker adapter — — — gemma nakasyou 0 17
Srizzle/triton-demo finetune — — — apache-2.0 Srizzle 0 0

Repository changelog

13 commits

Repository changelog 13 commits

DateCommitSHA
2025-12-16 Add readme 7c38f16
2025-12-15 Add readme 74f0553
2025-12-12 Add readme 63bcdbd
2025-12-12 Add readme 55527be
2025-12-12 Add readme 15634e1
2025-12-12 Add readme 84f760f
2025-12-12 Add readme 1126c5d
2025-12-11 Upload folder using huggingface_hub f45ffaf
2025-11-25 Upload folder using huggingface_hub f6246b4
2025-11-25 Upload folder using huggingface_hub 82957ed
2025-10-25 Upload folder using huggingface_hub 372c488
2025-10-25 Upload folder using huggingface_hub 750fa6a
2025-10-25 initial commit dc28ba8

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

T5Gemma 2: The next generation of encoder-decoder models
blog.google2025-12-18officialauthorityHN front pageabout the t5gemma 2 familyHN 153 pts · 28 comments619 words
T5Gemma 2 launches as the first multi-modal and long-context encoder-decoder models based on Gemma 3, featuring tied embeddings and merged attention mechanisms. The family offers compact pre-trained models at sizes of 270M-270M (~370M total), 1B-1B (~1.7B), and 4B-4B (~7B) parameters. These models support over 140 languages and handle context windows of up to 128K tokens.written, reviewed, cited
  • The release explicitly states that no post-trained or instruction-tuned checkpoints will be made available to the public.
  • Developers must perform their own supervised fine-tuning because the provided pre-trained checkpoints are designed specifically for post-training.
  • The model achieves its multimodal capabilities by utilizing a separate, highly efficient vision encoder alongside the text encoder.

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).

Similar models

3 models

Similar models

  • Florence-2-large · 777M
    same size band (S), same task (image-text-to-text), permissive license
  • Qwen3.5-0.8B · 873M
    same size band (S), same task (image-text-to-text), more providers, permissive license
  • Florence-2-large-ft · 770M
    same size band (S), same task (image-text-to-text), permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:42 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:13 UTC · Last verified 2026-10-25

Follow t5gemma-2-270m-270m

Watch this model

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