Gemma Model Comparison
Every Gemma model side by side: the five Gemma 4 sizes, Gemma 4 against earlier generations, the specialised variants, and how Gemma differs from Gemini. Pick any two and see exactly where they differ.
Figures from Google's model card and technical report ·
Compare any two models
Choose two models. The better figure on each row is shown in green.
| First model | Second model | |
|---|---|---|
| Turn on JavaScript to use the comparison tool, or see the full tables below. | ||
Scores are percentages from Google's model card and technical report, with thinking on for Gemma 4. Memory is Google's estimate for loading the weights; download sizes are from Ollama's tags page. Lower is better for memory and download size.
The five Gemma 4 sizes
Same family, same interface, very different hardware needs.
| Model | Design | Parameters | Context | Inputs | Memory, 4-bit | Memory, full |
|---|---|---|---|---|---|---|
| Gemma 4 31B | Dense | 30.7B | 256K | Text, image, video | 17.5 GB | 69.9 GB |
| Gemma 4 26B A4B | Mixture of Experts | 25.2B (3.8B active) | 256K | Text, image, video | 14.4 GB | 57.7 GB |
| Gemma 4 12B | Unified, encoder-free | 11.95B | 256K | Text, image, video, audio | 6.7 GB | 26.7 GB |
| Gemma 4 E4B | Edge | 4.5B effective | 128K | Text, image, video, audio | 4.5 GB | 17.9 GB |
| Gemma 4 E2B | Edge | 2.3B effective | 128K | Text, image, video, audio | 2.9 GB | 11.4 GB |
From Google's model card. Video is read as image frames on every size; audio only on 12B, E4B and E2B. All are Apache 2.0. File types for each size are on the models page; full figures in technical specs; memory planning in the VRAM guide and QAT builds.
Scores by size
The three larger models sit close together. The step down to the edge models is much bigger.
MMLU Pro
LiveCodeBench v6
| Test | 31B | 26B A4B | 12B | E4B | E2B |
|---|---|---|---|---|---|
| MMLU Pro · knowledge | 85.2 | 82.6 | 77.2 | 69.4 | 60.0 |
| AIME 2026 · maths | 89.2 | 88.3 | 77.5 | 42.5 | 37.5 |
| GPQA Diamond · science | 84.3 | 82.3 | 78.8 | 58.6 | 43.4 |
| LiveCodeBench v6 · coding | 80.0 | 77.1 | 72.0 | 52.0 | 44.0 |
| Tau2 · agent tasks | 76.9 | 68.2 | 69.0 | 42.2 | 24.5 |
| MMMU Pro · vision | 76.9 | 73.8 | 69.1 | 52.6 | 44.2 |
| RULER at 128K · long context | 96.4 | 89.8 | 91.2 | 86.6 | 70.4 |
Best in each row in green. Note the 12B edges past the 26B A4B on agent tasks and long-context retrieval. Every test, including audio and the Arena ranking, is on the benchmarks page; what the numbers mean in practice is under capabilities.
The four decisions people get stuck on
31B vs 26B A4B
The 31B is a few points ahead on most tests and clearly ahead on agent tasks and long context. The 26B A4B uses about 3.8B of its parameters per token, so it answers faster and needs 3.1 GB less memory. Choose the 26B for everyday speed, the 31B when quality decides the outcome.
MMLU Pro 85.2 vs 82.6 · coding 80.0 vs 77.1 · memory 17.5 vs 14.4 GB
26B A4B vs 12B
The 26B A4B is stronger on knowledge, maths and coding. The 12B needs less than half the memory, accepts audio, and matches or beats it on agent tasks and 128K retrieval. On a 16 GB machine the 12B is the answer.
MMLU Pro 82.6 vs 77.2 · coding 77.1 vs 72.0 · memory 14.4 vs 6.7 GB
12B vs E4B
This is the biggest drop in the family: about 20 points on coding and 35 on maths. Move to E4B only when the 12B will not fit, such as on a phone or an 8 GB laptop.
MMLU Pro 77.2 vs 69.4 · coding 72.0 vs 52.0 · memory 6.7 vs 4.5 GB
E4B vs E2B
E4B is better across the board. E2B exists for the tightest devices: Google lists a 2.58 GB phone build against 3.66 GB for E4B. See the Android page for measured speeds.
MMLU Pro 69.4 vs 60.0 · coding 52.0 vs 44.0 · memory 4.5 vs 2.9 GB
Verdicts are this site's reading of Google's figures. More on the 12B, phones on the Android page and mobile downloads.
Gemma 4 vs Gemma 3
A generational jump, not a refresh: a more open licence, double the context, audio and video input, a thinking mode and far higher scores.
| Gemma 4 | Gemma 3 | |
|---|---|---|
| Released | March 31, 2026 | March 2025 |
| Sizes | E2B, E4B, 12B, 26B A4B, 31B | 270M, 1B, 4B, 12B, 27B |
| Licence | Apache 2.0 | Gemma Terms of Use |
| Largest context | 256K | 128K |
| Inputs | Text, image, video, audio on three sizes | Text and image |
| Thinking mode | Yes | No |
| System role | Native | Folded into the user turn |
| Function calling | Built in | Through prompting |
| Test | Gemma 3 27B | Gemma 4 31B | Change |
|---|---|---|---|
| MMLU Pro | 67.6 | 85.2 | +17.6 |
| AIME 2026 | 20.8 | 89.2 | +68.4 |
| GPQA Diamond | 42.4 | 84.3 | +41.9 |
| LiveCodeBench v6 | 29.1 | 80.0 | +50.9 |
| Tau2 | 16.2 | 76.9 | +60.7 |
| MMMU Pro | 49.7 | 76.9 | +27.2 |
| RULER at 128K | 66.0 | 96.4 | +30.4 |
From Google's Gemma 4 model card and technical report. Gemma 3 has no thinking mode, so its scores are without it. Gemma 3 still offers one thing Gemma 4 does not: a tiny 270M model. More on the Gemma 3 page.
Every Gemma generation
Five releases in just over two years. Gemma 4 is the first under the Apache 2.0 licence.
| Generation | Released | Sizes | Largest context | Inputs | Licence |
|---|---|---|---|---|---|
| Gemma 4 | March 2026 | E2B, E4B, 12B, 26B A4B, 31B | 256K | Text, image, video, audio | Apache 2.0 |
| Gemma 3n | June 2025 | E2B, E4B | 32K | Text, image, video, audio | Gemma Terms |
| Gemma 3 | March 2025 | 270M, 1B, 4B, 12B, 27B | 128K | Text, image | Gemma Terms |
| Gemma 2 | June 2024 | 2B, 9B, 27B | 8K | Text | Gemma Terms |
| Gemma 1 | February 2024 | 2B, 7B | 8K | Text | Gemma Terms |
Earlier generations are summarised from Google's Gemma release history and were not re-checked line by line for this update; some sizes arrived after the first release date shown.
Specialised Gemma variants
Not bigger or smaller chat models - different tools for different jobs.
EmbeddingGemma 2 New · Oct 6, 2026
Turns text, code, images, video and audio into vectors for search and retrieval. 740M parameters, 768 dimensions, Apache 2.0. It does not generate text.
DiffusionGemma Experimental
A 26B A4B sibling that writes by refining blocks of text in parallel instead of one token at a time: faster output, lower benchmark scores. Aimed at in-line editing and code infilling.
Earlier variants, built on older Gemma generations
These predate Gemma 4 and use the older Gemma Terms of Use. Dates and sizes are in Google's release history. For general coding, plain Gemma 4 is now the stronger choice - see Gemma 4 coding.
Gemma vs Gemini
Same research lab, different products. Gemma is the open model you download and run. Gemini is the larger, closed model Google runs for you.
| Gemma 4 | Gemini | |
|---|---|---|
| What it is | Open-weight models you download | Proprietary models offered as a service |
| How you get it | Hugging Face, Kaggle, Ollama, or a hosted API | Gemini app, Gemini API, Google Cloud |
| Runs offline | Yes - phone, laptop or your own servers | No |
| Your data | Stays on your hardware when self-hosted | Processed in Google's cloud |
| Fine-tune and modify | Full access to the weights | No access to the weights |
| Cost | Free weights; you pay for compute | Per-token pricing or a subscription |
| Raw capability | Strong for its size | Higher - Google's frontier models |
| Upkeep | You run and scale it | Managed by Google |
A common setup is Gemma 4 locally for routine and private work, with the hardest requests sent to a hosted model. Costs are on the pricing page; hosted access is on the API page. Against other open models, see Gemma 4 vs Qwen 3.5; against a coding tool, Gemma 4 vs Claude Code.
Which Gemma should you pick?
| If you need | Pick | Why |
|---|---|---|
| The best answers | Gemma 4 31B | Top score on every test in the table above |
| Speed on a 24 GB GPU | Gemma 4 26B A4B | Close to the 31B while using a fraction of its weights per token |
| A 16 GB laptop | Gemma 4 12B | 6.7 GB at 4-bit, 256K context, and it accepts audio |
| Audio input | Gemma 4 12B, E4B or E2B | The two largest models do not take audio |
| A phone or offline device | Gemma 4 E2B or E4B | Built for on-device use |
| Search and retrieval | EmbeddingGemma 2 | Purpose-built for embeddings |
| Something under 1 GB | Gemma 3 270M | Gemma 4 has no model that small |
| The strongest model, with no setup | Gemini | Hosted, larger and managed for you |
Ready to get one? Use the download hub, how to download and install guide; choose a runtime with the runtime selector; Mac users see Mac setup; teams see self-hosting. Try before installing in the playground or chat.
Which Gemma 4 model is the best?
The 31B scores highest. Whether it is best for you depends on memory: the 26B A4B is close behind and faster, and the 12B is the strongest model that fits a 16 GB laptop.
Is the 31B worth it over the 26B A4B?
Sometimes. The gap is two to three points on knowledge, maths and coding, but wider on agent tasks (76.9 vs 68.2) and 128K retrieval (96.4 vs 89.8). If those matter, yes.
Why does a 26B model run faster than a 31B one?
The 26B A4B is a Mixture-of-Experts model. Each token uses about 3.8B of its parameters, so there is far less to compute per token. All of it still has to fit in memory.
Should I upgrade from Gemma 3?
For almost every use, yes. Scores are far higher, the licence is more permissive and you gain thinking, audio and a longer context. The exception is if you rely on the 270M model.
Does the Apache 2.0 licence apply to older Gemma models?
No. Gemma 3 and earlier remain under the Gemma Terms of Use. Only Gemma 4 and the models built on it are Apache 2.0.
What does the "E" in E2B and E4B mean?
"Effective". The edge models use techniques that let them run with the memory footprint of a 2B or 4B model while holding more parameters in total. See features.
Is Gemma 4 as good as Gemini?
No. Gemini's frontier models are larger and more capable. Gemma 4's case is that you can download it, run it privately and modify it. Known gaps are in limitations.
Are these comparisons independent tests?
No. The figures are Google's own published results, not measurements made by this site. Test on your own tasks before deciding; prompt advice is in the prompting guide and responsible-use notes under safety.
Found your model?
Download it, or look closer at the numbers first.