Gemma 4

Google DeepMind (Gemma Team) · 2026-07-24 · analysed 2026-08-10 · skill v0.4.0

dense efficient-attention multimodal open-weights large on-device

Design overview

Gemma 4 is the fourth generation of Google DeepMind's open Gemma family, a suite of five natively multimodal decoder-only Transformers spanning dense 31B, 12B, and the per-layer-embedding E4B/E2B models plus one sparse MoE (26B total, 3.8B active), all sharing one SentencePiece tokenizer with 262k entries. Architecturally it is a direct refinement of Gemma 3: the backbone keeps the 5:1 local sliding-window to global attention interleaving (4:1 on E2B) with dual RoPE bases (10k local, 1M global), but adds pp-RoPE with p=0.25 on global layers, reuses keys as values in global attention (values=keys, except E2B/E4B), and shares KV caches across layers at 20/35 or 18/42 ratios — together cutting the global KV cache by up to 37.5%. Blocks keep RMSNorm pre- and post-norm with QK-norm. Vision attaches through frozen ViT encoders (550M for 31B/26B-A4B, 150M for E2B/E4B) with axial 2D-RoPE and aspect-ratio-preserving resizing; audio exists only on E2B/E4B via a slimmed 305M USM-style Conformer encoder, while the flagship novelty is the 12B's encoder-free design, which feeds raw 40ms audio chunks and 48×48×3 image patches through a 35M matmul directly into the LLM. Every model ships with an MTP drafter for speculative decoding and QAT-quantized checkpoints (mobile int2/int4, Q4_0). A thinking mode — reasoning traces toggled by a <|think|> token — is the family's headline behavioral feature, and Arena results put the 31B at the top of open dense models.

What the report claims is novel

Atlas assessment: A conservative, efficiency-focused evolution of the google-gemma-3 skeleton: same dense GQA-class decoder with 5:1 local/global attention (4:1 on E2B), dual RoPE bases, RMSNorm pre/post-norm + QK-norm, and the Gemini/SentencePiece tokenizer. The genuinely new mechanisms for the atlas are incremental rather than structural: keys-as-values reuse on global layers (values=keys, from Kayyam et al. 2026), cross-layer KV cache sharing (Shazeer 2019 MQA lineage), pp-RoPE with p=0.25 partial rotation on global layers (Barbero et al. 2025), per-layer embeddings on the small dense models (Gemma 3n lineage, not an atlas record), an EAGLE-style MTP drafter (deepseek-v3's MTP is the nearest atlas peer), QAT-shipped quantized checkpoints, and o1-style thinking mode. The encoder-free 12B (raw audio chunks + image patches through a 35M matmul) is the most distinctive claim — no atlas peer ingests raw patches without an encoder. Notably, the report does not name GQA, the FFN activation, or any MoE routing detail for 26B-A4B, and audio input is absent from the 31B and 26B-A4B variants despite the 'natively multimodal' framing. No new attention or mixing family is introduced.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyGemma 4[1]
OrganisationGoogle DeepMind (Gemma Team)[2]
Release date2026-07-24[3]
Report URLhttps://arxiv.org/abs/2607.02770[4]
LicenseApache 2.0[5]
Open weightsyes[6]

Variants

Scale

Total parameters31.25B
Active parameters / token31.25B
Layersnot disclosed
Hidden dimensionnot disclosed
FFN inner dimensionnot disclosed
FFN rationot disclosed
Query headsnot disclosed
KV headsnot disclosed
Head dimensionnot disclosed
Vocabulary size262000[7]
Embedding tyingnot disclosed

Per-variant configs

VariantTotalLayersHiddenFFNQ headsKV headsContext
E2B5.065Bnot disclosednot disclosednot disclosednot disclosednot disclosednot disclosed
E4B7.885Bnot disclosednot disclosednot disclosednot disclosednot disclosednot disclosed
12B11.89Bnot disclosednot disclosednot disclosednot disclosednot disclosednot disclosed
26B-A4B25.79Bnot disclosednot disclosednot disclosednot disclosednot disclosednot disclosed
31B31.25Bnot disclosednot disclosednot disclosednot disclosednot disclosednot disclosed

Core block

Block typedense
Attention varianthybrid[8]
Attention layer pattern5:1 interleaving of local sliding-window attention layers per global self-attention layer for all models except E2B (4:1); local window span not disclosed; global layers reuse keys as values (values=keys) except in E2B and E4B; KV cache shared across layers at ratios 20/35 (E2B) and 18/42 (E4B)[9]
Depth mixingsequential-residual
Activationnot disclosed
Expert countn/a[10]
Experts per tokenn/a[10]
Shared expertsn/a[10]
Routingn/a[10]
Load balancingn/a[10]
Expert granularityn/a[10]
MethodRoPE[9]
RoPE base frequency1000000[11]
Partial RoPEyes[12]
TypeRMSNorm[13]
Placementmixed[14]
QK-normyes[15]
Attention sinksno
Softcappingno
OtherA scalar scale at each block bounds activation ranges to fit fp16 for stable inference (§2.5 QAT section)[16]

Context

Trained context lengthnot disclosed
Deployed context lengthnot disclosed
Extension methodnot disclosed

Tokenizer

AlgorithmSentencePiece[17]
NotesSame tokenizer as Gemini Team [2025]; 262k vocabulary; IT formatting adds control tokens (<|think|>, <|turn|>, <|tool_call|>, <|channel>) for thinking mode and function calling (Table 11)[18]

Training

Training tokensnot disclosed
Data compositionLarge-scale, diverse collection from a wide range of domains and modalities, including web documents, code, images, and audio (audio for E2B, E4B and 12B); data cutoff date January 2025; filtered to decontaminate benchmarks and reduce unwanted or unsafe utterances and recitation risk[19]
Curriculum / stagingnone disclosed[20]
Optimizernot disclosed
LR schedulenot disclosed
Batch schedulenot disclosed
Precisionnot disclosed
ParallelismZeRO-3 optimizer-state sharding; Pathways data-replica reduction over the data-center network; GSPMD partitioner with the MegaScale XLA compiler; Slice-Granularity Elasticity for failure recovery on the larger models; sharding by data, sequence, and replica (Table 2)[21]
HardwareTPUv4 and TPUv6e: E2B 4,096 v6e chips; E4B 6,144 v6e; 12B 12,288 v4; 26B-A4B 6,144 v6e; 31B 10,240 v6e[21]
Disclosed computenot disclosed

Post-training

SFTyes
Preference optimisationnone disclosed[22]
Reasoning trainingThinking mode: models generate a reasoning trace before responding (o1-style, cited OpenAI [2024]); activated via the <|think|> token in a leading system turn (Table 11); the training recipe behind the thinking mode is not disclosed[23]
Distillationnot disclosed

Modality

Typemultimodal[24]
Attachmentadapter[25]

Inference efficiency

KV-cache designLocal sliding-window layers bound KV growth to the window span; global layers reuse keys as values (values=keys, except E2B/E4B) and use pp-RoPE p=0.25, reducing the global KV cache footprint by up to 37.5%; KV caches shared across layers at ratios 20/35 (E2B) and 18/42 (E4B); int8 KV caching offered (Table 3, '+KV' at 32k context)[26]
Quantisation shippedQAT-trained quantized checkpoints: mobile quantization (per-channel low-bitwidth weights, int2/int4 mix, with int8 activation quantization) and Q4_0 blockwise quantization; raw checkpoints bf16; QAT also applied to encoders — 150M vision encoder W8A8 (forward memory 400→200 MB), audio encoder int8 activations with {2,4,8}-bit weights by layer cluster (on-disk 390→87 MB)[27]
Speculative decodingAutoregressive MTP drafter head for speculative decoding: a 4-layer Transformer block (d_model 256 for E2B/E4B, 1024 for 26B-A4B/31B; 3 local + 1 global attention layers) with a separate embedder, cross-attending to the main model's KVs — no MTP prefill, any draft length; E2B/E4B drafters use cluster top-k decoding (d×262,000 → d×4,096)[28]
Serving optimisationsnone disclosed[29]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
Arena Text Elo1451Table 4
MMLU Pro85.2Table 5
AIME 2026 (no tools)89.2Table 5
LiveCodeBench v680.0Table 5
Codeforces Elo2150Table 5
GPQA Diamond84.3Table 5
HLE19.5Table 5
MMMU Pro (1120 vision tokens)76.9Table 6
MATH-Vision (1120 vision tokens)85.6Table 6
InfographicVQA (1120 vision tokens)92.0Table 6
RULER accuracy 128k96.4Table 9
LOFT Text Retrieval Recall@k 128k79.5Table 9

Provenance references

  1. stated · Title
  2. stated · Title page byline
  3. stated · Header: 'arXiv:2607.02770v2 [cs.CL] 24 Jul 2026'
  4. stated · pipeline
  5. stated · §1 ('We release the Gemma 4 models under an Apache 2.0 license')
  6. stated · Abstract ('open-weight, natively multimodal language models'); §1 (Apache 2.0 release)
  7. stated · §2.4 Tokenizer ('The vocabulary has 262k entries'); Table 1 caption ('The vocabulary we use has 262k entries')
  8. stated · §1, §2 (Long-context efficiency: '5:1 ratio of local sliding window to global self-attention (4:1 for the 2.3B model)')
  9. stated · §2 (Long-context efficiency)
  10. n/a · §2 (Dense and MoE paragraph)
  11. stated · §2 ('The RoPE frequencies are set to 1M and 10k on global and local attention layers, respectively')
  12. stated · §2 ('We encode position with pp-RoPE with p=0.25 on global attention layers')
  13. stated · §2 ('Our models have pre-norm and post-norm with RMSNorm')
  14. stated · §2 ('pre-norm and post-norm with RMSNorm')
  15. stated · §2 ('and QKNorm [Henry et al., 2020]')
  16. stated · §2.5
  17. stated · §2.4 Tokenizer ('a SentencePiece tokenizer with split digits, preserved whitespace, and byte-level encodings')
  18. stated · §2.4, §3 (PT vs IT formatting), Table 11
  19. stated · §2.4
  20. stated · §2.4 ('We follow a similar pre-training as Gemma 3')
  21. stated · §2.7, Table 2
  22. stated · §3
  23. stated · §1, §3, Table 11
  24. stated · Abstract ('natively multimodal'); §1 ('capable of seamlessly processing text, images, and audio')
  25. stated · §1 ('Gemma 4 models have frozen vision and audio encoders'), §2.1
  26. stated · §1, §2, Table 3
  27. stated · §2.5, Table 3
  28. stated · §1, §2.6
  29. stated · Report-wide absence (no serving stack described)