AI Radar

Gemma 4

Available · Language, Multimodal, Reasoning · Milestone

Gemma 4 is a family of open-weight models from Google, released on 2 April 2026 under the Apache 2.0 licence in E2B, E4B, 26B Mixture-of-Experts and 31B dense sizes, with a 12B Unified size added in June 2026. All sizes take text, image and video input and return text. Google says Gemma 4 is built from the same research and technology as Gemini 3. [1] [2] [3] Primary source

Timeline of Gemma 4 →

Claims and evidence

  • Released Launch sizes E2B, E4B, 26B A4B and 31B; the 12B Unified size followed on 2026-06-03 (see variants).Other sources give: [4]March 31, 2026 The Gemma releases page dates the release of Gemma 4 in E2B, E4B, 31B and 26B A4B sizes to March 31, 2026. Primary source[1] [5]April 2, 2026
  • Status AvailableOpen weights for all sizes are published in Google's Hugging Face organisation (as seen on 2026-10-01). Primary source[6]
  • Successor of Gemma 3The technical report introduces Gemma 4 as building on its predecessors and cites the Gemma, Gemma 2 and Gemma 3 reports; the launch post refers only to the first generation. Primary source[7]1 Introduction
  • Change · Licensing Released under the Apache 2.0 licence, a choice Google links to developer feedback.Compared with Gemma 3 Primary source[1]
  • Change · Reasoning Adds a configurable thinking mode in all sizes, which produces a reasoning trace before the answer.Compared with Gemma 3 Primary source[7]1 Introduction [2]Core Capabilities: Thinking
  • Change · Modality Adds native audio input on the E2B and E4B sizes, and on the 12B Unified size added in June 2026.Compared with Gemma 3 Primary source[2]
  • Change · Architecture Adds a Mixture-of-Experts size, 26B A4B, which activates 3.8 billion of its 25.2 billion parameters during inference.Compared with Gemma 3 Primary source[2]Mixture-of-Experts (MoE) Model
  • Input text, image, videoAll sizes take image and video input; E2B, E4B and 12B also take audio (see variants). Primary source[1]Vision and audio [2]introduction: Extended Multimodalities
  • Output text Primary source[2]introduction
  • Feature Reasoning modeConfigurable thinking mode in all sizes. Primary source[2]Core Capabilities: Thinking [7]1 Introduction
  • Feature Function calling Primary source[1]Agentic workflows [2]Core Capabilities: Function Calling
  • Open weights YesReleased under the Apache 2.0 licence, pre-trained and instruction-tuned checkpoints. Primary source[6] [1]An ecosystem of choices: Download the models
  • Context window 256K tokensLarger sizes (12B, 26B A4B, 31B); E2B and E4B have 128K (see variants). Primary source[1]Longer context [2]Core Capabilities: Long Context
  • Access Open-weights download, API, On deviceWeights on Hugging Face, Kaggle and Ollama; 26B A4B and 31B also in Google AI Studio and the Gemini API; E2B and E4B run offline on phones and edge devices (Google AI Edge Gallery, AICore Developer Preview). Primary source[1]An ecosystem of choices; E2B and E4B models [5]April 2, 2026

Lineage

Predecessors

Successors

No known successor.

Based on

Not derived from another model.

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

None.

Variants

Gemma 4 E2B

Same dates as Gemma 4.

  • Variant Inline variant in this record.Edge size with per-layer embeddings; pre-trained and instruction-tuned checkpoints. Primary source[1]We are releasing Gemma 4 in four versatile sizes [6]

Also known as: gemma-4-E2B-it

Differs in:

  • Input: text, image, audio, video Evidence not assessed [1]Vision and audio [2]introduction: Extended Multimodalities; Dense Models: Supported Modalities
  • Context window: 128K tokens Evidence not assessed [2]Dense Models: Context Length
  • Parameters: 2.3B effective (5.1B with embeddings) Evidence not assessed [2]Dense Models: Total Parameters

Gemma 4 E4B

Same dates as Gemma 4.

  • Variant Inline variant in this record.Edge size with per-layer embeddings; pre-trained and instruction-tuned checkpoints. Primary source[1]We are releasing Gemma 4 in four versatile sizes [6]

Also known as: gemma-4-E4B-it

Differs in:

  • Input: text, image, audio, video Evidence not assessed [1]Vision and audio [2]introduction: Extended Multimodalities; Dense Models: Supported Modalities
  • Context window: 128K tokens Evidence not assessed [2]Dense Models: Context Length
  • Parameters: 4.5B effective (8B with embeddings) Evidence not assessed [2]Dense Models: Total Parameters

Gemma 4 26B A4B

Same dates as Gemma 4.

  • Variant Inline variant in this record.Mixture-of-Experts size (8 of 128 experts active plus 1 shared); the launch post calls it 26B Mixture of Experts (MoE). Also served through the Gemini API as gemma-4-26b-a4b-it. Primary source[1]26B and 31B models [6] [5]April 2, 2026

Also known as: gemma-4-26b-a4b-it

Differs in:

  • Parameters: 25.2B total, 3.8B active Evidence not assessed [2]Mixture-of-Experts (MoE) Model

Gemma 4 31B

Same dates as Gemma 4.

  • Variant Inline variant in this record.Dense size; the launch post calls it 31B Dense. Also served through the Gemini API as gemma-4-31b-it. Primary source[1]26B and 31B models [6] [5]April 2, 2026

Also known as: gemma-4-31b-it

Differs in:

  • Parameters: 30.7B Evidence not assessed [2]Dense Models: Total Parameters

Gemma 4 12B Unified

  • Released Primary source[3] [4]June 3, 2026
  • Variant Inline variant in this record.Added after the launch; encoder-free architecture that feeds image patches and audio directly into the language model. The announcement headline calls it Gemma 4 12B; the releases page and model card call it Gemma 4 12B Unified. Primary source[3] [6]

Also known as: Gemma 4 12B, gemma-4-12B-it

Differs in:

  • Input: text, image, audio, video Evidence not assessed [3]first paragraph; Novel unified architecture [2]introduction: Extended Multimodalities; Dense Models: Supported Modalities
  • Parameters: 11.95B Evidence not assessed [2]Dense Models: Total Parameters

Related AI Radar coverage

All model releases from Google on AI Radar →