Gemma 4 Technical Report (Google DeepMind, July 2026)

The Gemma 4 Technical Report is a paper by the Gemma Team at Google DeepMind describing the fourth generation of the Gemma open-weight family (Source: https://arxiv.org/abs/2607.02770). The arXiv HTML captured here is version 2, dated July 24, 2026; the models themselves were announced on April 2, 2026. The report covers five instruction-tuned models — E2B and E4B (2.3B and 4.5B effective parameters, with per-layer embeddings), a 12B, a 26B Mixture-of-Experts with 3.8B active parameters (26B-A4B), and a 31B dense model — and describes them as natively multimodal, with audio input on the E2B, E4B and 12B models. It is the document that the shorter Gemma 4 model card cites for its benchmark table.

Architecture and training

The report lists five changes from Gemma 3. A thinking mode lets the model emit a reasoning trace before answering. For long context, the models keep a 5:1 ratio of local sliding-window to global attention layers (4:1 for E2B), use proportional RoPE, share KV cache and reuse keys as values in global layers; context reaches 256K tokens for the 12B, 26B-A4B and 31B and 128K for the E-series. A multi-token-prediction drafter head is released for speculative decoding. Quantization-aware training provides mobile (int2/int4 weight, int8 activation) and Q4_0 checkpoints, and the report states that quantizing the 150M image encoder to W8A8 halves its forward-pass memory from 400 MB to 200 MB. The 12B model uses an encoder-free design that projects raw 40 ms audio chunks and image patches directly into the embedding space through linear layers; the other models use frozen vision (about 150M or 550M parameters) and audio (about 305M) encoders. Table 1 gives parameter counts by component. Pre-training uses the Gemini tokenizer with a 262k vocabulary and TPUv4 and TPUv6e; post-training follows the Gemma 3 recipe with the addition of thinking, and filters personal information, unsafe outputs, mistaken self-identification and duplicates.

Evaluation approach

Section 4 reports human evaluation on the Arena Text leaderboard and static benchmarks. Table 5 compares all five Gemma 4 models, in thinking mode, with Gemma 3 27B in non-thinking mode across MMLU Pro, AIME 2026 (no tools), LiveCodeBench v6, Codeforces Elo, SciCode, GPQA Diamond, Big Bench Extra Hard, Humanity's Last Exam (with and without search), IFBench, IFEval, MMMLU, MRCR v2 (8-needle, 128k), Terminal Bench Hard and the three τ²-bench domains. Table 6 gives vision results at the maximal 1,120-token resolution, with Table 12 repeating them at 280 tokens. Tables 7 and 8 report CoVoST speech-translation BLEU and FLEURS ASR error rates by language for the E2B, E4B and 12B models against Gemma 3n. Table 9 reports long-context results (RULER, LOFT text retrieval, GraphWalks and MTOB) without thinking.

Findings

On the Arena Text leaderboard as of June 19, 2026, the report lists Gemma 4 31B at 1451 Elo (rank 43) and 26B-A4B at 1438 (rank 61), against Gemma 3 27B at 1366 (rank 157), and includes Claude Fable 5 at 1508 as the top closed model "for scale"; it describes Gemma 4 31B as "the leading dense open model on the leaderboard" and says both models match much larger open MoE models such as GLM 5.x, Kimi K2.x and DeepSeek V4. On Table 5, Gemma 4 31B scores 85.2 on MMLU Pro, 89.2 on AIME 2026, 80.0 on LiveCodeBench v6, 2150 Codeforces Elo, 84.3 on GPQA Diamond, 19.5 on HLE without tools (26.5 with search), 66.4 on MRCR v2 8-needle at 128k and 36.0 on Terminal Bench Hard; the authors state that the 31B is "significantly better across the board" than Gemma 3 27B and that E2B "roughly matches" Gemma 3 27B with a tenth of the parameters. On vision, the report says E4B equals or outperforms Gemma 3 27B on every listed benchmark. On audio, the caption to Table 7 reports relative improvements over Gemma 3n of 12% (E2B) and 10% (E4B) on translation and 17% and 12% on transcription, with the audio encoder's on-disk footprint reduced from 390 MB to 87 MB.

Responsibility, safety and security

Section 5 describes governance, safety policies and train-time mitigations, safety evaluations, ethical considerations and the approach to open release. The safety-evaluation subsection states that automated and human evaluations, run without safety filters, showed "major improvements in every category of content safety relative to previous Gemma models", that the models "significantly outperform Gemma 3 and 3n" while keeping unjustified refusals low, and that policy violations were "minimal" for text-to-text and image-to-text across sizes. It refers to the commitments in Google's Frontier Safety Framework. No violation rates, refusal rates or dangerous-capability scores 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 arXiv HTML (v2, 24 July 2026) as captured on September 21, 2026. All figures are reported by Google DeepMind; Arena comparison rows are Google's transcription of the public leaderboard on the stated date. Confidence is medium: a single developer-authored source with fast-decay benchmark claims.

Relationships

• supports: Gemma (Google open-weight models)
• supersedes: Gemma 3 Technical Report (Google DeepMind, March 2025) (as the current-generation report; the earlier report remains the record for Gemma 3)
• instance-of: AI System Cards
• related: Gemma 4 Model Card (Google DeepMind, 2026), Google DeepMind, LMArena (Chatbot Arena), AIME (as a model benchmark), Humanity's Last Exam, τ-bench, MRCR