Gemma is Google DeepMind's family of open-weight models, built from the same research as Gemini, and its fourth generation, Gemma 4, was released on 31 March 2026 under the Apache 2.0 licence. This hub lists every Gemma model in Google's public Gemma documentation as of 15 November 2026, with sizes, parameter counts, context windows, modalities, licences, API IDs, prices, availability and benchmark figures. It covers Gemma 4 and the earlier Gemma 1, 2, 3 and 3n generations, plus the specialised variants Google lists. It is written for developers, buyers, architects and anyone who needs to know which Gemma model is which. Models are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from Google's own documentation, cited in the sources list.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Google's public Gemma documentation as of November 2026. Model details change, so confirm against Google's current documentation.
Key takeaways
- Gemma 4 has five sizes. Google lists E2B, E4B, 12B Unified, 26B A4B and 31B; the first four arrived on 31 March 2026 and the 12B Unified model on 3 June 2026.
- Context is 128K or 256K tokens. The E2B and E4B models list 128K; the 12B, 26B A4B and 31B models list 256K.
- Gemma 4 moved to Apache 2.0. The Gemma 4 model card states Apache 2.0, while Gemma 1 to Gemma 3n remain under the Gemma Terms of Use, last modified on 1 April 2026.
- The Gemini API offers Gemma at no charge. gemma-4-31b-it and gemma-4-26b-a4b-it are free on the Gemini API free tier, with no paid tier listed; free-tier data is used to improve Google's products.
- Google publishes model cards, not system cards. Each generation has a model card with benchmarks, safety evaluation and limitations, and Gemma 4 has a technical report.
What is Google Gemma?
Gemma is a family of open models from Google DeepMind that Google describes as lightweight and built from the same technology that powers its Gemini models. Unlike Gemini, which is offered only as a hosted service, Gemma weights can be downloaded from Kaggle and Hugging Face and run on a developer's own hardware, from phones and browsers to cloud GPUs and TPUs. Google also serves the two largest Gemma 4 models through the Gemini API and Google AI Studio. For Google's hosted models, see Google Gemini Models.
Gemma has had five main generations: Gemma (February 2024), Gemma 2 (June 2024), Gemma 3 (March 2025), Gemma 3n (June 2025, on-device) and Gemma 4 (March 2026). Google also publishes variants built on these bases for embeddings, vision, safety classification, function calling, translation and medical use. Google's ai.google.dev Gemma pages now carry a notice that they are deprecated in favour of new documentation at gemma.ai; both were checked for this page. Facts are as of 15 November 2026.
Current Gemma models at a glance
As of 2026-11-15. Source: Gemma 4 model card, Gemma models overview, gemma.ai models page and the Gemini API Gemma page. Models are listed by release date, newest first. Category is Google's architecture label.
| Model | API ID or version | Category | Context | Licence | Status |
|---|---|---|---|---|---|
| Gemma 4 12B Unified | Open weights (no Gemini API ID stated) | Unified, encoder-free multimodal | 256K | Apache 2.0 | Current |
| Gemma 4 31B | gemma-4-31b-it | Dense | 256K | Apache 2.0 | Current |
| Gemma 4 26B A4B | gemma-4-26b-a4b-it | Mixture of experts | 256K | Apache 2.0 | Current |
| Gemma 4 E4B | Open weights (no Gemini API ID stated) | Small, effective parameters | 128K | Apache 2.0 | Current |
| Gemma 4 E2B | Open weights (no Gemini API ID stated) | Small, effective parameters | 128K | Apache 2.0 | Current |
| Gemma 3n E2B, E4B | gemma-3n-e4b-it (preview on the Gemini API) | On-device | Not stated by Google on the pages checked (2026-11-15) | Gemma Terms of Use | Earlier generation |
| Gemma 3 (270M, 1B, 4B, 12B, 27B) | gemma-3-27b-it (Gemini API) | Dense | 128K (4B, 12B, 27B); 32K (1B, 270M) | Gemma Terms of Use | Earlier generation |
Every Gemma 4 size supports configurable thinking modes, built-in function calling and a native system role, according to the models overview. Multi-token prediction (MTP) draft models for speculative decoding were released on 16 April 2026 for E2B, E4B, 31B and 26B A4B. See Open-Weight vs Closed Models Explained.
Key facts for the main Gemma models
As of 2026-11-15. Identical fields for every model; "Not stated by Google" means not stated by Google on the pages checked (2026-11-15). Parameter counts are those in the model card.
| Model | Released | Max output | Knowledge cutoff | Input / output | Open weights | Parameters | Positioning (provider wording) |
|---|---|---|---|---|---|---|---|
| Gemma 4 12B Unified | 2026-06-03 | Not stated by Google | January 2025 | Text, image, audio, video as frames / text | Yes | 11.95B, 48 layers | "Encoder free" model that projects image patches and audio directly into the decoder |
| Gemma 4 31B | 2026-03-31 (weights); 2026-04-02 (Gemini API) | Not stated by Google | January 2025 | Text, image, video as frames / text | Yes | 30.7B, 60 layers | Dense model |
| Gemma 4 26B A4B | 2026-03-31 (weights); 2026-04-02 (Gemini API) | Not stated by Google | January 2025 | Text, image, video as frames / text | Yes | 25.2B total, 3.8B active; 128 experts, 8 active plus 1 shared | Mixture-of-experts model; all 26B parameters must be loaded into memory |
| Gemma 4 E4B | 2026-03-31 | Not stated by Google | January 2025 | Text, image, audio, video as frames / text | Yes | 8B with embeddings, 4.5B effective | Small model for edge and browser deployment |
| Gemma 4 E2B | 2026-03-31 | Not stated by Google | January 2025 | Text, image, audio, video as frames / text | Yes | 5.1B with embeddings, 2.3B effective | "Ultra-mobile, edge, and browser deployment" |
| Gemma 3 27B | 2025-03-10 | Remaining context after input | August 2024 | Text, image / text | Yes | 27B (14T training tokens) | Largest Gemma 3 size |
All Gemma 4 sizes share a 262K-token vocabulary. Sliding-window attention uses 512 tokens on E2B and E4B and 1,024 on the larger sizes. The "E" in E2B and E4B stands for effective parameters: Per-Layer Embeddings make the total count, and the memory needed, larger than the effective figure. See Mixture of Experts Explained.
Memory needed to load Gemma 4 weights
Approximate GPU or TPU memory from Google's models overview, including 20% overhead and excluding KV cache and context memory. Fine-tuning needs much more.
| Model | BF16 | SFP8 | Q4_0 | Mobile (LiteRT-LM) |
|---|---|---|---|---|
| Gemma 4 31B | 69.9 GB | 34.9 GB | 17.5 GB | Not stated by Google |
| Gemma 4 26B A4B | 57.7 GB | 28.8 GB | 14.4 GB | Not stated by Google |
| Gemma 4 12B Unified | 26.7 GB | 13.4 GB | 6.7 GB | Not stated by Google |
| Gemma 4 E4B | 17.9 GB | 8.9 GB | 4.5 GB | 2.5 GB (2.2 GB text only) |
| Gemma 4 E2B | 11.4 GB | 5.7 GB | 2.9 GB | 1.1 GB (0.84 GB text only) |
Google also publishes quantisation-aware training (QAT) checkpoints in GGUF, compressed-tensors and mobile formats, which it states perform nearly identically to the high-precision baselines.
Gemma model families and naming
- Gemma 4: the current generation, with text, image and video input at every size and audio on E2B, E4B and 12B.
- Gemma 3n: on-device models in E2B and E4B sizes with text, image and audio input.
- Gemma 3: 270M, 1B, 4B, 12B and 27B; text only at 270M and 1B, text and image from 4B up.
- Gemma 2 and Gemma 1: text-only models in 2B, 9B and 27B (Gemma 2) and 2B and 7B (Gemma 1) sizes.
- Specialised variants: CodeGemma, RecurrentGemma, PaliGemma and PaliGemma 2 (vision-language), ShieldGemma and ShieldGemma 2 (safety classifiers), DataGemma, Gemma-APS, T5Gemma, VaultGemma, MedGemma, TranslateGemma, FunctionGemma (270M), EmbeddingGemma (308M) and EmbeddingGemma 2 (740M, multimodal, built on Gemma 4). gemma.ai also lists DiffusionGemma, an experimental text diffusion model on Gemma 4, without a release date.
- Size labels: "31B" is the parameter count; "26B A4B" means 26B total with about 4B active per token; "E2B" means about 2B effective parameters.
- Suffixes: "-it" marks instruction-tuned checkpoints; pre-trained checkpoints are published alongside. See LLM Naming and Versioning Explained.
When was each Gemma model released?
As of 2026-11-15. Dates are from Google's Gemma releases page and the Gemini API change log. Developer guides and tooling entries are omitted.
| Date | Event | API ID |
|---|---|---|
| 2026-10-06 | EmbeddingGemma 2 (740M) | Open weights |
| 2026-06-03 | Gemma 4 12B Unified | Open weights |
| 2026-04-16 | Gemma 4 MTP draft models for E2B, E4B, 31B and 26B A4B | Open weights |
| 2026-04-02 | Gemma 4 31B and 26B A4B released in AI Studio and the Gemini API | gemma-4-31b-it, gemma-4-26b-a4b-it |
| 2026-03-31 | Gemma 4 (E2B, E4B, 31B, 26B A4B) | Open weights |
| 2026-01-15 | TranslateGemma (4B, 12B, 27B) | Open weights |
| 2026-01-13 | MedGemma 1.5 (4B) | Open weights |
| 2025-12-18 | FunctionGemma (270M); T5Gemma 2 | Open weights |
| 2025-09-13 | VaultGemma (1B) | Open weights |
| 2025-09-04 | EmbeddingGemma (300M on the releases page) | Open weights |
| 2025-08-14 | Gemma 3 270M | Open weights |
| 2025-06-26 | Gemma 3n (E2B, E4B) | Open weights |
| 2025-05-20 | gemma-3n-e4b-it preview on the Gemini API; MedGemma (4B, 27B) | gemma-3n-e4b-it |
| 2025-03-12 | Gemma 3 27B on the Gemini API | gemma-3-27b-it |
| 2025-03-10 | Gemma 3 (1B, 4B, 12B, 27B); ShieldGemma 2 | Open weights |
| 2024-12-05 | PaliGemma 2 (3B, 10B, 28B) | Open weights |
| 2024-06-27 | Gemma 2 (9B, 27B); Gemma 2 2B followed on 2024-07-31 | Open weights |
| 2024-04-09 | CodeGemma and RecurrentGemma | Open weights |
| 2024-02-21 | Gemma (2B, 7B) | Open weights |
The latest model release on the releases page is EmbeddingGemma 2 on 6 October 2026; the latest general-purpose model is Gemma 4 12B Unified on 3 June 2026. See the LLM Release Tracker.
How much do Gemma models cost?
As of 2026-11-15. Price per million tokens in USD on the Gemini API. Google lists Gemma 4 as one entry without per-size prices. Source: Gemini API pricing page.
| Model | Input | Cached input | Output | Batch | Cache write | Notes |
|---|---|---|---|---|---|---|
| Gemma 4 (Gemini API, free tier) | Free of charge | Free of charge | Free of charge | Not stated by Google | Context caching storage free of charge | Data used to improve Google's products: yes |
| Gemma 4 (Gemini API, paid tier) | Not available | Not available | Not available | Not stated by Google | Not available | Paid tier listed as not available |
| Open weights (all generations) | No licence fee | N/A | No licence fee | N/A | N/A | Hardware and hosting costs fall on the operator |
- Self-hosting: downloading Gemma weights from Kaggle or Hugging Face carries no fee from Google; compute, storage and operation are paid by whoever runs the model.
- Google Cloud: Google Cloud documentation lists a managed Model-as-a-Service (MaaS) offering for Gemma-4-26B-A4B-IT and self-deployment from Model Garden; MaaS prices are not stated on the Gemma pages checked (2026-11-15).
- Rate limits: not stated on the Gemini API Gemma page checked.
See Model Pricing Explained and the LLM Pricing Tracker.
Where are Gemma models available?
- Downloads: Kaggle and the Google Gemma 4 collection on Hugging Face.
- Google-hosted: Google AI Studio and the Gemini API (gemma-4-31b-it, gemma-4-26b-a4b-it), with thinking, image understanding, system instructions, function calling and Google Search grounding.
- Google Cloud: Gemini Enterprise Agent Platform documentation describes Model Garden deployment on GPU or TPU and a MaaS endpoint for Gemma-4-26B-A4B-IT; the models overview also lists Google Cloud and GKE.
- On-device: Google AI Edge and LiteRT-LM for mobile and browser deployment.
- Local runtimes and frameworks named by Google: llama.cpp, MLX, LM Studio, Ollama, Hugging Face Transformers, Keras, vLLM, SGLang, Tunix and Unsloth.
Self-hosting moves operation, patching and access control to the organisation. See Running LLMs Locally Explained and the On-Device AI Security Guide.
Restricted-access Gemma models
As of 2026-11-15. Source: Gemma models overview, Gemini API Gemma page and pricing page.
| Model | Access stated by provider | Data retention condition |
|---|---|---|
| Gemma 4 (all sizes, weights) | Public download from Kaggle and Hugging Face | Not applicable to self-hosted use; data stays with the operator |
| gemma-4-31b-it, gemma-4-26b-a4b-it (Gemini API) | Public API with an API key from Google AI Studio; free tier only | Free-tier data is used to improve Google's products |
| Gemma 1 to Gemma 3n and variants | Public download subject to the Gemma Terms of Use and Prohibited Use Policy | Not applicable to self-hosted use |
No Gemma model is described as restricted to verified or invited organisations on the pages checked.
What is the context window of Gemma models?
Gemma 4 E2B and E4B list a 128K-token context window, and Gemma 4 12B Unified, 26B A4B and 31B list 256K. Gemma 3 lists 128K for the 4B, 12B and 27B models and 32K for the 1B and 270M models, with output limited to whatever context remains after the input. The Gemma 2 model card checked does not state a context length. Google does not state a separate maximum output length for Gemma 4. Running long contexts locally needs extra memory for the KV cache beyond the weight figures above. See Context Windows Explained.
Is any Gemma model open source?
All Gemma models are open-weight. The Gemma 4 model card states the Apache 2.0 licence, a standard open-source licence that permits commercial use, modification and redistribution. Gemma 1, 2, 3 and 3n and their variants are governed by the Gemma Terms of Use (last modified 1 April 2026), which Google states permit responsible commercial use but which add obligations: distributors must pass on the use restrictions in Section 3.2, give recipients a copy of the terms, mark modified files and include a Notice file, and all users must follow the Gemma Prohibited Use Policy. Training data is not published; the Gemma 4 card describes web documents in over 140 languages, code, mathematics, images and audio. See the Open-Weight Model Security Guide.
Model cards and system cards
As of 2026-11-15.
| Model | Model card or page | System card | Technical report |
|---|---|---|---|
| Gemma 4 | Gemma 4 model card (ai.google.dev) | Not stated by Google on the pages checked (2026-11-15) | Gemma 4 technical report (linked from the models overview) |
| Gemma 3 and 3n | Gemma 3 and Gemma 3n model cards | Not stated by Google on the pages checked (2026-11-15) | Not stated by Google on the pages checked (2026-11-15) |
| Gemma 2 | Gemma 2 model card | Not stated by Google on the pages checked (2026-11-15) | Not stated by Google on the pages checked (2026-11-15) |
The Gemma 4 card states that safety testing ran without safety filters, covered child sexual abuse material, dangerous content, sexually explicit content, hate speech and harassment, and found minimal policy violations in text-to-text and image-to-text output across all sizes. It lists limitations including training data gaps, missed nuance and factual errors, and states that training data was filtered for CSAM and certain personal and sensitive information. See Model Cards and System Cards Explained.
Benchmarks
As of 2026-11-15. Provider-reported and independent results are kept in separate columns. Scores are instruction-tuned results as Google reports them in the Gemma 4 and Gemma 3 model cards.
| Model | Provider-reported benchmarks | Independent leaderboard results |
|---|---|---|
| Gemma 4 31B | MMLU Pro 85.2%; AIME 2026 (no tools) 89.2%; LiveCodeBench v6 80.0%; GPQA Diamond 84.3%; MMMU Pro 76.9% | None cited on this page |
| Gemma 4 26B A4B | MMLU Pro 82.6%; AIME 2026 88.3%; LiveCodeBench v6 77.1%; GPQA Diamond 82.3%; MMMU Pro 73.8% | None cited on this page |
| Gemma 4 12B Unified | MMLU Pro 77.2%; AIME 2026 77.5%; LiveCodeBench v6 72.0%; GPQA Diamond 78.8%; MMMU Pro 69.1% | None cited on this page |
| Gemma 4 E4B | MMLU Pro 69.4%; AIME 2026 42.5%; LiveCodeBench v6 52.0%; GPQA Diamond 58.6%; MMMU Pro 52.6% | None cited on this page |
| Gemma 4 E2B | MMLU Pro 60.0%; AIME 2026 37.5%; LiveCodeBench v6 44.0%; GPQA Diamond 43.4%; MMMU Pro 44.2% | None cited on this page |
| Gemma 3 27B | MMLU-Pro 67.5; GPQA Diamond 42.4; LiveCodeBench 29.7 (Gemma 3 card, 0-shot) | None cited on this page |
The Gemma 4 card also lists Gemma 3 27B without thinking at MMLU Pro 67.6% and LiveCodeBench v6 29.1%, slightly different from the Gemma 3 card because benchmark versions and settings differ. See How to Read LLM Benchmarks.
Deprecations and retirements
As of 2026-11-15. Source: Gemini API change log and deprecations page, Gemma releases page.
| Model | API ID | Deprecated or notice | Retirement or shutdown | Stated replacement |
|---|---|---|---|---|
| Gemma 3 27B (Gemini API) | gemma-3-27b-it | Not stated by Google on the pages checked (2026-11-15) | Not stated by Google on the pages checked (2026-11-15) | Not stated by Google |
| Gemma 3n E4B (Gemini API preview) | gemma-3n-e4b-it | Not stated by Google on the pages checked (2026-11-15) | Not stated by Google on the pages checked (2026-11-15) | Not stated by Google |
| Open-weight checkpoints (all generations) | N/A | N/A | Downloaded weights do not expire | N/A |
| ai.google.dev Gemma documentation | N/A | Marked deprecated in favour of gemma.ai | Not stated by Google | gemma.ai |
Hosted Gemini API model IDs can be retired by Google, while self-hosted weights stay usable for as long as the operator runs them, along with any vulnerabilities the operator does not fix. See the LLM Deprecation and Retirement Tracker.
Data handling and security
- Self-hosted use: prompts and outputs stay in the operator's environment; Google receives no inference data from downloaded weights.
- Gemini API free tier: the pricing page states that free-tier content is used to improve Google's products; the Gemma paid tier is listed as not available.
- Outputs: the Gemma Terms of Use state that Google claims no rights in outputs and that users are responsible for them.
- Use restrictions: Gemma 1 to 3n carry the Prohibited Use Policy, and Google may restrict use it reasonably believes breaches the terms.
- Safety classifiers: ShieldGemma and ShieldGemma 2 are published for text and image safety classification.
- Supply chain: pull weights from Google's official Kaggle and Hugging Face accounts and verify what you deploy. See the Self-Hosted LLM Security Guide.
See AI Data Retention and Deletion and the Foundation Model Selection Guide.
How to choose a Gemma model: factual criteria
- Hardware: compare the memory table: the 31B needs 69.9 GB in BF16 or 17.5 GB at 4-bit; E2B runs in about 1.1 GB on mobile.
- Modalities: audio input is listed only on E2B, E4B and 12B Unified; all Gemma 4 sizes take text, image and video.
- Context: 256K on 12B, 26B A4B and 31B; 128K on E2B and E4B.
- Active compute: 26B A4B activates 3.8B parameters per token but needs all weights loaded.
- Hosting model: the Gemini API offers 31B and 26B A4B on the free tier with product-improvement data use; self-hosting keeps data in-house.
- Licence: Apache 2.0 for Gemma 4; the Gemma Terms of Use for earlier generations and variants.
- Specialist need: embeddings, translation, medical, safety classification and function calling each have a dedicated variant.
Practitioner checklist
- Record the exact Gemma generation, size and checkpoint (pre-trained, instruction-tuned or QAT) in the AI inventory.
- Confirm which licence applies: Apache 2.0 for Gemma 4, the Gemma Terms of Use and Prohibited Use Policy for earlier models.
- Size hardware from Google's memory table plus KV cache for the context length you plan to use.
- Do not send confidential data through the Gemini API free tier, where content is used to improve Google's products.
- Download weights only from Google's official Kaggle and Hugging Face accounts and record hashes.
- Patch and update self-hosted inference servers and runtimes on a schedule.
- Store Gemini API keys in a secrets manager and rotate them. See the LLMjacking Guide.
Standards and references
- Google: Gemma 4 model card (accessed 2026-11-15)
- Google: Gemma models overview (accessed 2026-11-15)
- Google: Gemma releases (accessed 2026-11-15)
- Google DeepMind: Gemma models (accessed 2026-11-15)
- Google: Gemini API release notes (accessed 2026-11-15)
- Google: Gemini API pricing (accessed 2026-11-15)
- Google: Run Gemma with the Gemini API (accessed 2026-11-15)
- Google: Gemma Terms of Use (accessed 2026-11-15)
- Google: Gemma 3 model card (accessed 2026-11-15)
Related NHI Mgmt Group resources: Google Gemini Models, Small and Low-Cost LLMs Compared, Self-Hosted LLMs vs Hosted LLM APIs, Open-Weight Model Security Guide.