Gemma 3 Technical Report
Summary
Gemma 3 is a multimodal addition to the Gemma family of lightweight open models, ranging from 1 to 27 billion parameters. This version introduces vision understanding, wider language coverage, and longer context of at least 128K tokens. The authors change the architecture to reduce KV-cache memory under long context by increasing the ratio of local to global attention layers and keeping the local attention span short. They train the models with distillation and report superior performance to Gemma 2 for both pretrained and instruction-finetuned versions, including that Gemma3-4B-IT is competitive with Gemma2-27B-IT. They release all of the models to the community. [S1]
Claims
- id: C1 claim: 'Gemma 3 is a multimodal Gemma family ranging from 1 to 27 billion parameters.' source: S1 evidence: "lightweight open models, ranging in scale from 1 to 27 billion parameters"
- id: C2 claim: 'The authors report longer context of at least 128K tokens.' source: S1 evidence: "longer context - at least 128K tokens"
- id: C3 claim: 'They say they release all of the models to the community.' source: S1 evidence: "We release all our models to the community"
Verification
- Code available: No official model or training-code URL was named on the abs page or HTML, even though the authors say they release all of the models to the community. No URL was checked as this paper's primary release.
- Reproduced: not reproduced (2026-11-12). Code not run.
- Related work: Live note 2403-05530-gemini-1-5.
- Limits: Company technical report (v1, 25 Mar 2025). Comparisons with Gemma 2 and Gemini 1.5 Pro are author-reported. The paper does not name a model-hosting URL.
- License: CC BY 4.0 on the abs page (2026-11-12; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 25 Mar 2025 (v1). Sources: [S1] https://arxiv.org/abs/2503.19786, v1 (25 Mar 2025). checked_on: 2026-11-12.
When to apply
Use when asking what the Gemma 3 report claimed for multimodal open models and 128K-token context.
Facts
- id: F1 claim: 'Gemma 3 is a multimodal open-model family ranging from 1 to 27 billion parameters.' confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F2 claim: 'The authors report context of at least 128K tokens and an architecture change aimed at reducing KV-cache memory.' confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F3 claim: 'They say they release all of the models to the community.' confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
Sources
- [S1] arXiv 2503.19786, https://arxiv.org/abs/2503.19786. License: CC BY 4.0. v1 (25 Mar 2025). Accessed 2026-11-12.
License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.
License: CC-BY-4.0 · Summary by iinventedlight (AI-assisted). Source: Gemma Team and 215 coauthors (listed on the arXiv abs page), Gemma 3 Technical Report (CC BY 4.0).
Cite this
'iinventedlight (AI-assisted)'. 2026. "Gemma 3 Technical Report." iinventedlight Library, iil:note:qzakqbxa, version of 2026-11-12. https://iinventedlight.com/library/notes/2503-19786-gemma-3/. License: CC-BY-4.0.
@misc{iil_note_qzakqbxa,
author = {{'iinventedlight (AI-assisted)'}},
title = {{Gemma 3 Technical Report}},
howpublished = {iinventedlight Library},
year = {2026},
date = {2026-11-12},
url = {https://iinventedlight.com/library/notes/2503-19786-gemma-3/},
note = {Permanent ID iil:note:qzakqbxa. License: CC-BY-4.0.},
}
Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.
For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2503-19786-gemma-3.md.