AI Radar

Gemma 3

Available · Language, Multimodal · Milestone

Gemma 3 is a family of open-weight models from Google, released on 12 March 2025 in 1B, 4B, 12B and 27B sizes, with a 270M size added in August 2025. Google says it is built from the same research and technology as Gemini 2.0; the 4B, 12B and 27B sizes accept images as well as text and have a 128K-token context window. [1] [2] [3] [4]Table 1 [5]Inputs and outputs Primary source

Timeline of Gemma 3 →

Claims and evidence

  • Released Other sources give: [6]March 10, 2025 Date given on the Gemma releases page. Primary source[2] [1] [7]March 12, 2025
  • Status AvailableWeights remain downloadable from Google's Hugging Face organisation (checked 2026-10-01). Primary source[8] [5]Resources and Technical Documentation
  • Successor of Gemma 2The technical report presents Gemma 3 as the newest version of the Gemma models and compares it with Gemma 2 throughout. Primary source[4]Abstract; section 1
  • Change · Modality Adds image input for the 4B, 12B and 27B sizes through a shared vision encoder; the 1B size stays text-only.Compared with Gemma 2 Primary source[4]Table 1; section 2, vision encoder
  • Change · Context length Context window raised to 128K tokens for the 4B, 12B and 27B sizes and to 32K for the 1B size.Compared with Gemma 2 Primary source[4]section 2, long context
  • Change · Languages Wider language coverage: out-of-the-box support for over 35 languages and pretrained support for over 140.Compared with Gemma 2 Primary source[2]New capabilities
  • Change · Architecture Uses five local attention layers per global layer, with a short local span to limit KV-cache memory, and QK-norm instead of the soft-capping of Gemma 2.Compared with Gemma 2 Primary source[4]
  • Input text, imageImage input applies to the 4B, 12B and 27B sizes, which share one SigLIP-based vision encoder; the 1B size is text-only (technical report, Table 1). The announcement also mentions short videos. Primary source[5]Inputs and outputs [1]Multimodality
  • Output text Primary source[5]Inputs and outputs
  • Feature Function calling Primary source[2]New capabilities [1]
  • Feature MultilingualOut-of-the-box support for over 35 languages and pretrained support for over 140. Primary source[2]New capabilities
  • Open weights YesPre-trained and instruction-tuned weights under the Gemma Terms of Use; access on Hugging Face requires accepting the licence. Primary source[8] [2]Get started with Gemma 3
  • Context window 128K tokensFor the 4B, 12B and 27B sizes; the 1B and 270M sizes have 32K. Primary source[2]New capabilities [5]Inputs and outputs
  • Access Open-weights download, APIWeights on Hugging Face and Kaggle; gemma-3-27b-it in Google AI Studio and the Gemini API; also offered through Vertex AI Model Garden. Primary source[2]Get started with Gemma 3 [7]March 12, 2025

Lineage

Predecessors

Successors

Based on

Not derived from another model.

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

Variants

Gemma 3 1B

Same dates as Gemma 3.

  • Variant Inline variant in this record. Primary source[2]introduction (sizes) [4]Table 1

Also known as: gemma-3-1b-it

Differs in:

  • Input: text Evidence not assessed [4]Table 1 (no vision encoder for 1B); section 2, vision encoder
  • Context window: 32K tokens Evidence not assessed [4]section 2, long context [5]Inputs and outputs
  • Parameters: 1B Evidence not assessed [2]introduction (sizes) [4]Table 1

Gemma 3 4B

Same dates as Gemma 3.

  • Variant Inline variant in this record. Primary source[2]introduction (sizes) [4]Table 1

Also known as: gemma-3-4b-it

Differs in:

  • Parameters: 4B Evidence not assessed [2]introduction (sizes) [4]Table 1

Gemma 3 12B

Same dates as Gemma 3.

  • Variant Inline variant in this record. Primary source[2]introduction (sizes) [4]Table 1

Also known as: gemma-3-12b-it

Differs in:

  • Parameters: 12B Evidence not assessed [2]introduction (sizes) [4]Table 1

Gemma 3 27B

Same dates as Gemma 3.

  • Variant Inline variant in this record. Primary source[2]introduction (sizes) [4]Table 1 [8]

Also known as: gemma-3-27b-it

Differs in:

  • Parameters: 27B Evidence not assessed [2]introduction (sizes) [4]Table 1

Gemma 3 270M

  • Released Primary source[3] [6]August 14, 2025
  • Variant Inline variant in this record.Added later as a compact size for task-specific fine-tuning, with pre-trained and instruction-tuned checkpoints. Primary source[3]

Also known as: gemma-3-270m

Differs in:

  • Context window: 32K tokens Evidence not assessed [5]Inputs and outputs
  • Parameters: 270 million Evidence not assessed [3]Core capabilities

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Google on AI Radar →