Gemma 4 Model Card (Google DeepMind, 2026)

The Gemma 4 model card is the page on Google's AI developer site that documents the fourth generation of the Gemma open-weight family from Google DeepMind (Source: https://ai.google.dev/gemma/docs/core/model_card_4). The captured version carries a "Last updated 2026-07-30" stamp; the family was announced on April 2, 2026. The card links the Hugging Face collection, GitHub, the launch blog and the Gemma 4 Technical Report (arXiv 2607.02770), states that the models are released under the Apache 2.0 licence, and lists five sizes: E2B, E4B, 12B, 26B A4B and 31B.

Models overview

Two tables give per-model properties. The dense models are E2B (2.3B effective parameters, 5.1B with embeddings, 35 layers, 128K context, text/image/audio), E4B (4.5B effective, 8B with embeddings, 42 layers, 128K, text/image/audio), 12B Unified (11.95B, 48 layers, 256K, text/image/audio) and 31B Dense (30.7B, 60 layers, 256K, text/image). The MoE model, 26B A4B, has 25.2B total and 3.8B active parameters, 30 layers, 8 active of 128 experts plus one shared, 256K context and text/image input. The card explains that "E" denotes effective parameters made possible by per-layer embeddings, that "A" in 26B A4B denotes active parameters, and that "Unified" refers to the 12B's encoder-free architecture, which projects raw image patches and audio waveforms into the embedding space through lightweight linear layers. All models use a hybrid of local sliding-window and global attention with the final layer always global, unified keys and values in global layers, and proportional RoPE.

Benchmark results

A single table, described as covering instruction-tuned models, reports the five Gemma 4 models beside "Gemma 3 27B (no think)" on MMLU Pro, AIME 2026 (no tools), LiveCodeBench v6, Codeforces Elo, GPQA Diamond, Tau2 (average over three domains), Humanity's Last Exam (no tools, and with search, for the larger models), BigBench Extra Hard, MMMLU, MMMU Pro, OmniDocBench 1.5 (edit distance, lower is better), MATH-Vision, MedXPertQA MM, CoVoST and FLEURS (audio models only; 12B values exclude Chinese), and MRCR v2 8-needle at 128k. The card does not state whether the Gemma 4 rows were run in thinking mode; the technical report's fuller version of the same table says they were. Gemma 4 31B is the highest-scoring model in every row.

Capabilities, usage and data

The card lists thinking, long context, image understanding (including OCR, document parsing, UI understanding and pointing), video via frame sequences, interleaved multimodal input, native function calling, coding, 35+ supported languages with pre-training on 140+, and audio (ASR and speech-to-translated-text) on E2B, E4B and 12B. A best-practices section gives sampling parameters, thinking-mode configuration, modality ordering, variable image resolution and audio and video length limits. The training-data section describes a mix of web text, code, mathematics, images and audio, with filtering for CSAM, sensitive data and quality; the usage section lists intended uses, limitations and ethical considerations in general terms.

Ethics and safety

The card states that Gemma 4 "undergoes the same rigorous safety evaluations as our proprietary Gemini models", that automated and human evaluations were run against Google's safety policies (child sexual abuse material and exploitation, dangerous content, sexually explicit content, hate speech, harassment), and that testing was conducted without safety filters. It reports "major improvements in all categories of content safety relative to previous Gemma models", that Gemma 4 "significantly outperform[s] Gemma 3 and 3n models" while keeping unjustified refusals low, and "minimal policy violations" for text-to-text and image-to-text across sizes. No rates are given.

Reported results

Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.

Provenance

The page summarises the HTML card as captured on September 21, 2026. All figures are reported by Google DeepMind; the Gemma 3 27B comparison column is Google's own earlier model. The benchmark table is a subset of Table 5 of the technical report, which is the fuller record. Confidence is medium: a single developer-authored source with fast-decay benchmark claims.

Relationships

• supports: Gemma (Google open-weight models)
• depends-on: Gemma 4 Technical Report (Google DeepMind, July 2026)
• instance-of: AI System Cards
• related: Google DeepMind, MMLU-Pro, GPQA, AIME (as a model benchmark), MRCR