Summary
Google DeepMind released Gemma 3 on March 12, 2025 in 1B, 4B, 12B, and 27B sizes. The 1B model accepted text only and had a 32K-token context window; the 4B, 12B, and 27B models accepted text and images with 128K context. The release used Google's Gemma terms rather than Apache 2.0 or MIT.
Editorial correction, October 26, 2026
The model card assigns 128K context to 4B/12B/27B and 32K to 1B, while Google's developer explanation identifies 1B as text-only. The original summary and body incorrectly generalized vision and 128K context to every launch variant. Later additions to the Gemma family are not backdated to this four-size launch.
Revision record with the prior summary and metadata.
The original narrative below is preserved as published; use the revised summary and this note for the corrected account.
What Happened
On March 12, 2025, Google DeepMind published Gemma 3 in four sizes: 1B, 4B, 12B, and 27B parameters. All variants incorporated multimodal vision through a frozen SigLIP encoder, allowing image understanding without the full cost of end-to-end vision training. Context length extended to 128K tokens, and the models supported over 140 languages — a substantial improvement in multilingual coverage over earlier Gemma generations.
Benchmark results for the 27B model showed performance surpassing Gemini 1.5 Pro on multiple standard evaluations, positioning Gemma 3 as competitive with models that Google itself had previously charged commercial API rates for. The 1B model was designed specifically for on-device deployment on mobile hardware.
The release used the Gemma Terms of Use rather than a conventional open-source license. Notably, the terms prohibited using Gemma outputs to train models that compete with Google products — a restriction absent from Apache 2.0 and MIT licenses. This clause triggered criticism from open-source advocates and reignited a broader argument about whether models with commercial restriction carve-outs deserve the "open" label.
Why It Matters
Gemma 3 demonstrated that a lab releasing an open-weight model that exceeds its own prior-generation commercial product — while imposing distillation restrictions — occupies an ambiguous position in the open-vs-closed spectrum. The release sharpened the conceptual distinction between "weights available for download" and "fully open" in the tradition of open-source software.
It also raised a recurring structural question: when a proprietary lab releases open weights with usage restrictions, is this genuinely contributing to the open ecosystem, or is it a controlled release designed to capture goodwill while preserving competitive moats? The Gemma 3 license debate became a reference point in subsequent discussions of AI openness standards.
§ How to read the metadata
- Landmark
- Fundamentally alters the trajectory; 2–5 per year.
- Major
- Meaningfully shifts the landscape; 2–4 per month.
- Notable
- Worth documenting; significance can be upgraded later.
- Confidence
- High = primary sources corroborate. Medium = credible secondary only. Low = provisional. Disputed = credible sources disagree.
- Contestation
- Uncontested = no formal challenge. Contested = at least one challenge open. Superseded = replaced by a later entry. Unresolved = dispute still open.