Gemma 4 12B

Sits between Gemma 4 E4B and the 26B MoE, and is Google's first mid-sized Gemma with native audio input. Drops the separate multimodal encoders: vision goes through a lightweight embedding module and raw audio is projected straight into the text token space, so both flow into the LLM backbone directly.

Details

Benchmarks

Each row is a dated claim by a named party, not a property of the model. A later revision adds a row rather than replacing one. This project publishes no composite or overall score.

Release timeline

A model has more than one date. The timeline above positions it at its announcement; these are every recorded lifecycle event, each with the source that evidences it.

1. June 3, 2026 Announced blog.google, huggingface.co
2. June 3, 2026 Weights published blog.google

Sources

Record status: verified · confidence 90/100 · 2 of 2 primary

Cited to the lab's own announcement and the official model card. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities read from the launch announcement, which states: “Novel unified architecture: No multimodal encoders. The vision and audio inputs flow directly into the LLM backbone,” and describes each path separately — vision through a lightweight embedding module, raw audio projected into the same dimensional space as text tokens. Output is recorded as text: no image, audio or video generation is claimed anywhere in the primary sources. Capabilities evidenced from the primary sources: agentic: “5X 2X Today, we are introducing Gemma 4 12B, our latest model designed to bring agentic multimodal intelligence directly to laptops.” | vision: “Gemma 4 models are multimodal, handling text and image input (with audio supported on E2B, E4B, and 12B) and generating text output.” | coding: “Featuring both Dense and Mixture-of-Experts (MoE) architectures, Gemma 4 is well-suited for tasks like text generation, coding, and reasoning.” | reasoning: “Thinking Mode Configuration Compared to Gemma 3, the models use standard system , assistant , and user roles.”.

• blog.google Official announcement Primary archived
• huggingface.co Model card Primary archived

Primary means published by the organisation that made the model. Verified records require at least one.

Gemma family

1. Gemma February 21, 2024
2. Gemma 3 March 12, 2025
3. Gemma 4 12B June 3, 2026
4. EmbeddingGemma 2 October 6, 2026

What changed from Gemma 3

Compared with the previous release in this series. A field appears only when both records state a value — where one does not, it is listed as uncomparable rather than dropped, because a gap in our research is not a finding about the model.

Browse models

Compare Gemma 4 12B with another model →
See Gemma 4 12B on the interactive timeline →