Gemma 4 Specs
The complete datasheet for Google's open models: every size's parameters, layers, context, inputs and memory, how the models are built, Google's published benchmark results and on-device speeds. EmbeddingGemma 2 has its own spec sheet further down. Every figure names its source.
Spec viewer
Pick a size to see its full sheet.
Memory figures cover the model only; a conversation needs more on top. To compare two sizes side by side use the comparison tool; to choose a download see model variants.
Model specifications
| Specification | 31B | 26B A4B | 12B | E4B | E2B |
|---|---|---|---|---|---|
| Parameters | 30.7B | 25.2B 3.8B active | 11.95B | 4.5B eff. 8B total | 2.3B eff. 5.1B total |
| Layers | 60 | 30 | 48 | 42 | 35 |
| Context tokens | 256K | 256K | 256K | 128K | 128K |
| Local window tokens | 1,024 | 1,024 | 1,024 | 512 | 512 |
| Images | Yes | Yes | Yes | Yes | Yes |
| Video as frames | Yes | Yes | Yes | Yes | Yes |
| Audio | - | - | Yes | Yes | Yes |
| Vision part | 550M encoder | 550M encoder | 35M projection | 150M encoder | 150M encoder |
| Audio part | - | - | Raw projection | 305M encoder | 305M encoder |
Tokens, shared by all sizes.
Tokens an image; 280 by default.
Maximum, on the three sizes that take audio.
Maximum, at one frame a second.
From Google's Gemma 4 model card. "Effective" parameters for the edge models leave out their large embedding tables. Recommended sampling is temperature 1.0, top-p 0.95, top-k 64; see the prompt guide.
Memory by precision
How much memory the model itself needs at each level of compression.
| Precision | 31B | 26B A4B | 12B | E4B | E2B |
|---|---|---|---|---|---|
| Full precision BF16 | 69.9 GB | 57.7 GB | 26.7 GB | 17.9 GB | 11.4 GB |
| 8-bit | 34.9 GB | 28.8 GB | 13.4 GB | 8.9 GB | 5.7 GB |
| 4-bit Q4_0 | 17.5 GB | 14.4 GB | 6.7 GB | 4.5 GB | 2.9 GB |
| Mobile format | - | - | - | 2.5 GB | 1.1 GB |
Google's figures, from the model card. Official 4-bit builds trained for that format are explained on the QAT page; matching a size to your hardware is covered in the VRAM guide.
How the models are built
| Part | Detail |
|---|---|
| Type | Decoder-only transformer. Dense in four sizes; mixture of experts in the 26B A4B. |
| Attention pattern | Local and global layers interleaved at 5:1 (4:1 in E2B). Local layers look back 512 tokens in the edge models and 1,024 in the others. |
| Position encoding | p-RoPE with p = 0.25 on global layers, at a base frequency of 1M; standard RoPE at 10k on local layers. |
| Cache saving | Global layers reuse keys as values, cutting the global key-value cache by 37.5%. |
| Cache sharing, edge models | Later layers share cached keys and values: from layer 20 of 35 in E2B and layer 18 of 42 in E4B. |
| Vision | A 550M encoder in the 31B and 26B A4B; a 150M encoder in E4B and E2B; a 35M projection over 48 x 48 patches in the 12B. |
| Image detail | 70, 140, 280, 560 or 1,120 tokens an image; 280 by default. |
| Audio | A 305M encoder in E4B and E2B (25 tokens a second); raw audio in 40 ms frames at 16 kHz in the 12B. Clips of up to 30 seconds. |
| Video | Read as frames, up to 60 seconds at one frame a second. |
| Drafter models | Small companion models that speed up generation: 4 layers, with widths of 256 or 1,024 and 4,096 token clusters. The 12B drafter has about 0.4B parameters. |
| Vocabulary | About 262,000 tokens, shared across all sizes. |
| Training hardware | Google TPUv4 and TPUv6e. |
From Google's Gemma 4 technical report and model card. A plain-language walk through these ideas is on the features page.
Benchmark results
Google's published scores, with Gemma 3 27B for comparison. Higher is better unless a row says otherwise.
Text, code and agents
| Test | 31B | 26B A4B | 12B | E4B | E2B | Gemma 3 27B |
|---|---|---|---|---|---|---|
| MMLU Pro general knowledge | 85.2 | 82.6 | 77.2 | 69.4 | 60.0 | 67.6 |
| AIME 2026 competition maths | 89.2 | 88.3 | 77.5 | 42.5 | 37.5 | 20.8 |
| GPQA Diamond graduate science | 84.3 | 82.3 | 78.8 | 58.6 | 43.4 | 42.4 |
| BigBench Extra Hard hard reasoning | 74.4 | 64.8 | 53.0 | 33.1 | 21.9 | 19.3 |
| LiveCodeBench v6 coding | 80.0 | 77.1 | 72.0 | 52.0 | 44.0 | 29.1 |
| Codeforces coding, Elo rating | 2150 | 1718 | 1659 | 940 | 633 | 110 |
| Tau2 agent tasks | 76.9 | 68.2 | 69.0 | 42.2 | 24.5 | 16.2 |
| IFEval following instructions | 98.9 | 98.5 | 97.2 | 96.7 | 94.6 | 90.4 |
| RULER 128K long documents | 96.4 | 89.8 | 91.2 | 86.6 | 70.4 | 66.0 |
Images and documents
| Test | 31B | 26B A4B | 12B | E4B | E2B | Gemma 3 27B |
|---|---|---|---|---|---|---|
| MMMU Pro image understanding | 76.9 | 73.8 | 69.1 | 52.6 | 44.2 | 49.7 |
| MATH-Vision maths in images | 85.6 | 82.4 | 79.7 | 59.5 | 52.4 | 46.0 |
| MedXpertQA MM medical images | 61.3 | 58.1 | 48.7 | 28.7 | 23.5 | - |
| OmniDocBench 1.5 document reading; lower is better | 0.131 | 0.149 | 0.164 | 0.181 | 0.290 | 0.365 |
Audio
| Test | 31B | 26B A4B | 12B | E4B | E2B | Gemma 3 27B |
|---|---|---|---|---|---|---|
| FLEURS speech to text, word error rate; lower is better | - | - | 0.063 | 0.075 | 0.090 | - |
| CoVoST speech translation; higher is better | - | - | 38.5* | 35.54 | 33.47 | - |
Human preference
On the Arena text leaderboard on October 2, 2026 the 31B scored 1453 and the 26B A4B 1437, against 1365 for Gemma 3 27B.
With search
On Humanity's Last Exam the 31B scores 19.5 alone and 26.5 when it can search.
What these are
The maker's own numbers. This site has not re-run them. Gemma 3 27B was tested without thinking.
From Google's model card; Arena scores from lmarena.ai. *The 12B CoVoST figure excludes Chinese. Charts and discussion are on the benchmarks page; weak spots under limitations.
Speed on phones and laptops
Google's LiteRT-LM results for the two edge models.
| Device | Model | Runs on | Tokens a second | Time to first word | Memory |
|---|---|---|---|---|---|
| Galaxy S26 Ultra | E2B | Graphics chip | 52.1 | 0.3 s | 676 MB |
| Galaxy S26 Ultra | E2B | Processor | 46.9 | 1.8 s | 1,733 MB |
| Galaxy S26 Ultra | E4B | Graphics chip | 22.1 | 0.8 s | 710 MB |
| iPhone 17 Pro | E2B | Graphics chip | 56.5 | 0.3 s | 1,450 MB |
| iPhone 17 Pro | E4B | Graphics chip | 25.1 | 0.9 s | 3,380 MB |
| MacBook Pro M4 Max | E2B | Graphics chip | 160.2 | 0.1 s | - |
| MacBook Pro M4 Max | E4B | Graphics chip | 101.1 | 0.4 s | - |
From Google's LiteRT-LM pages for the E2B and E4B builds. Phone builds are 2.58 GB and 3.66 GB downloads with up to 32K tokens of context. Setup is on the Android page, under mobile downloads and on the Mac page.
EmbeddingGemma 2 specs
EmbeddingGemma 2 is a different kind of model. It does not write; it turns text, code, images, audio and video into lists of numbers, called vectors, that software can compare to find related items. It is built on the Gemma 4 architecture.
| Specification | Value |
|---|---|
| Parameters | 740M in full: 270M text and code, plus 170M vision and 300M audio |
| Configurations | 270M (text), 440M (text + vision), 570M (text + audio), 740M (everything) |
| Layers | 24 |
| Vector length | 768 numbers; can be cut to 512, 256 or 128 |
| Input budget | 8,192 tokens, shared by all input types |
| Fits in one input | About 29 images, 58 video frames or 327 seconds of audio |
| Inputs | Text, code, images, audio, video |
| Languages | 100+ |
| Training data cutoff | January 2025 |
| Licence | Apache 2.0 |
| Released | October 6, 2026 |
| Precision | bfloat16 or float32; not float16 |
| Library | sentence-transformers 6.1.0 or newer |
| Memory on a phone | About 191 MB text-only, 567 MB in full (Pixel 11 Pro) |
| Downloads | Ollama 378 MB to 1.3 GB; phone builds 165, 388 and 485 MB |
| Vector length | Text MTEB v2 | Images MIEB lite | Visual docs MMEB v2 |
|---|---|---|---|
| 768 | 61.36 | 64.64 | 59.01 |
| 512 | 61.17 | 64.32 | 58.38 |
| 256 | 60.41 | 63.13 | 56.24 |
| 128 | 57.89 | 59.06 | 45.65 |
MTEB Code. Version 1 scored 68.76.
Per million full-length vectors; 250 MB at 128.
From Google's EmbeddingGemma 2 model card, developer guide and on-device post. Scores are the maker's own; the model is days old and has not been independently tested. Usage is in how to use; storage costs on the pricing page.
Where each number comes from
| Figures | Source |
|---|---|
| Parameters, layers, context, memory, benchmarks | Gemma 4 model card, Google |
| Architecture detail | Gemma 4 technical report, Google DeepMind |
| QAT file sizes | Gemma 4 QAT collection on Hugging Face |
| Ollama names and sizes | Ollama's tag list |
| On-device speeds | Google's LiteRT-LM pages on Hugging Face |
| Human-preference scores | Arena text leaderboard, October 2, 2026 |
| EmbeddingGemma 2 | Model card, developer guide and on-device post, Google |
How many parameters does Gemma 4 have?
It depends on the size: 30.7B (31B), 25.2B (26B A4B), 11.95B (12B), 4.5B effective (E4B) and 2.3B effective (E2B).
What is Gemma 4's context window?
256K tokens on the 31B, 26B A4B and 12B; 128K on E4B and E2B. Phone builds are limited to 32K.
How much memory does Gemma 4 need?
From 2.9 GB (E2B) to 17.5 GB (31B) at 4-bit, for the model itself. At full precision, 11.4 GB to 69.9 GB.
What does "A4B" mean?
About four billion active parameters. The 26B A4B has 25.2B in total but uses roughly 3.8B for each token.
What does "effective" mean for E2B and E4B?
The count without their large embedding tables. With them, E2B has 5.1B parameters and E4B 8B.
Which sizes accept audio?
E2B, E4B and 12B, in clips of up to 30 seconds. See capabilities.
What licence is it under?
Apache 2.0, for all five sizes and for EmbeddingGemma 2. See ethics and safety for Google's use policies.
What are EmbeddingGemma 2's specs in short?
740M parameters, 24 layers, 768-number vectors, an 8,192-token input budget, five input types and 100+ languages.
Found the size you need?
Download it, or compare two sizes side by side.