Google Gemma

Google Gemma is Google DeepMind’s open-weights model family, built on the same research as Gemini. Unlike Gemini, Gemma models can be downloaded, run locally, fine-tuned and used commercially. The current models are the Gemma 4 language models (E2B to 31B, Apache 2.0) and the multimodal embedding model EmbeddingGemma 2 (since 6 October 2026), plus the experimental DiffusionGemma for fast local text generation. Because the weights are open, self-hosting keeps every piece of information inside your infrastructure.

innFactory AI Consulting, based in Rosenheim, Germany, advises enterprises across the DACH region and Europe on GDPR-compliant self-hosting of Google Gemma, from model selection through deployment to fine-tuning. This overview reflects the state as of October 2026.

Current models at a glance

Release dates per Google (Gemma release page, Google blog), parameters and context per the model cards on Hugging Face. Older generations (Gemma 3, Gemma 2) remain available for download, see previous generations. For the medical domain there is MedGemma.

Gemma 4: the current language model generation

Gemma 4 builds on the research behind Gemini 3 and, compared with Gemma 3, adds an integrated thinking mode, longer context windows and audio input on the smaller models. Google introduced the family on 2 April 2026 and released it under Apache 2.0.

Key characteristics

• Thinking mode: integrated reasoning for math, science and multi-step tasks
• Context: 256K tokens on 12B, 26B A4B and 31B, 128K tokens on E2B and E4B
• Multimodal: all models process text, images and video (as frames, up to about 60 seconds at 1 fps); audio input up to 30 seconds is supported only by E2B, E4B and 12B
• MoE variant: Gemma 4 26B A4B uses 128 experts, 8 of them active plus one shared, and computes with 3.8B parameters per token
• Effective parameters: E2B and E4B use per-layer embeddings, so the effective parameter count is far below the total
• Languages: out-of-the-box support for 35+ languages, pre-trained on 140+
• Agentic workflows: function calling and structured outputs for multi-step flows
• Knowledge: pre-training data up to January 2025 per the model card

Gemma 4 12B Unified, QAT and MTP

• Gemma 4 12B Unified (3 June 2026): mid-sized model with an encoder-free architecture. Image patches and audio waveforms are projected directly into the language model’s embedding space through lightweight linear layers instead of separate encoders. Per Google this reduces multimodal latency and allows the whole model to be fine-tuned in a single pass. With 256K context and about 12B parameters it runs, per Google, on laptops with 16 GB RAM/VRAM.
• QAT checkpoints (5 June 2026): quantization-aware training versions of all Gemma 4 sizes on Hugging Face, with up to 72% lower VRAM needs at near-identical quality to the BF16 original. Useful for consumer GPUs and on-device use.
• MTP checkpoints (16 April 2026): multi-token prediction variants for E2B, E4B, 26B A4B and 31B.

Benchmarks of the Gemma 4 family

Figures per the model card on Hugging Face (instruction-tuned):

The gap to Gemma 3 27B is largest in reasoning and code: LiveCodeBench rises from 29.1% to 80.0%, AIME 2026 from 20.8% to 89.2%. Gemma 4 12B Unified scores above Gemma 3 27B on every listed benchmark with less than half the parameters.

EmbeddingGemma 2: multimodal embeddings for local search and RAG

On 6 October 2026, Google DeepMind released EmbeddingGemma 2, the successor to the text-only EmbeddingGemma from 2025 (308M parameters). The model maps text (including code), images, video and audio into one shared vector space. A text query can therefore find matching images, video scenes or audio recordings without transcribing or captioning them first. It is released under Apache 2.0 and, per the model card, builds on Gemma 4.

Key facts

• Size: 740M parameters, built modularly from a backbone (130M), embedder (140M), vision encoder (170M) and audio encoder (300M). Unused encoders can be switched off: 270M for text, 440M for text and image, 570M for text and audio, 740M in full.
• Memory: per Google, quantized on a Pixel 11 Pro about 191 MB of RAM for text-only workloads and 567 MB for the fully multimodal model.
• Context: 8,192 tokens shared across all modalities, four times the predecessor. Per Google this covers up to 5.5 minutes of audio, 29 images or 58 video frames.
• Dimensions: 768, truncatable via Matryoshka representation learning to 512, 256 or 128 (up to six times less storage in the vector index). Per the model card, quality stays close to lossless down to 256 dimensions; 128 dimensions degrade multimodal tasks substantially.
• Languages: over 100, with quality varying by language.
• Knowledge: pre-training data up to January 2025.

Benchmarks per the model card

Figures at 768 dimensions and full precision (model card on Hugging Face). On text, quality stays at the predecessor’s level; on code the model gains about ten points.

Use and integration

• Typical applications: semantic search across image, video and audio archives, finding scenes in videos by text or audio query, on-device RAG together with Gemma 4, indexing local codebases for coding agents, classification and routing.
• Where to get it: weights on Hugging Face and Kaggle; supported frameworks include transformers, sentence-transformers, vLLM, SGLang, MLX, llama.cpp (GGUF), Ollama and LM Studio, transformers.js for the browser and MediaPipe and LiteRT for mobile devices. Google lists availability in the Gemini Enterprise Agent Platform Model Garden as ‘coming soon’; availability there is not yet documented.
• Practical notes: text inputs need short task prefixes (separate ones for query and document in search and Q&A), documents use the format title: {title} | text: {content}. Compute in bfloat16 or float32; per the model card, float16 can produce unusable embeddings without raising an error. The model has no safety post-training, so filters belong in the retrieval pipeline.

How it differs from Gemini Embedding 2: Google’s API model gemini-embedding-2 has been stable since 22 April 2026, returns up to 3,072 dimensions and also accepts PDF. It runs only as a Google service, though (see Google Gemini). EmbeddingGemma 2 is the choice when embeddings should be created locally, offline or entirely within your own infrastructure.

Specialized variants

DiffusionGemma (10 June 2026)

• Type: experimental open model using text diffusion, based on Gemma 4 26B A4B (25.2B parameters, 3.8B active), Apache 2.0
• Principle: instead of token by token, the model generates blocks of 256 tokens in parallel and refines them step by step; input text, image and video (no audio), context up to 256K
• Speed: per Google up to four times faster generation on dedicated GPUs, more than 1,000 tokens/s on one H100 and more than 700 tokens/s on an RTX 5090; quantized, it fits in 18 GB of VRAM
• Limits: per Google the advantage applies to local use with few concurrent requests; in high-load cloud serving it shrinks and can raise costs. Quality is below Gemma 4: MMLU Pro 77.6% (Gemma 4 26B A4B: 82.6%), LiveCodeBench v6 69.1% (77.1%), AIME 2026 69.1% (88.3%) per the model card
• Use: experiments with latency-critical local workflows such as in-line editing or code infilling; for production quality Google recommends standard Gemma 4

TranslateGemma (15 January 2026)

• Sizes: 4B, 12B and 27B parameters
• Focus: translation across 55 languages (per Google)
• Use: multilingual corporate communication, document localization without an external translation API

FunctionGemma (December 2025)

• Size: 270M parameters
• Focus: foundation for your own specialized function-calling models; per Google designed to be highly performant after fine-tuning
• Use: API integration, workflow automation, agents on small hardware

Gemma Scope 2 (December 2025)

• Type: interpretability suite for Gemma 3
• Benefit: insight into internal representations, for example for reviews in regulated industries

Previous generations: Gemma 3 and Gemma 2

Gemma 3 and Gemma 2 remain available for download and run in many existing deployments. They are covered by the Gemma Terms of Use, not Apache 2.0.

Gemma 3 Family (2025)

Gemma 2 Family (2024)

Comparison: Gemma 4 vs. Gemini 4 Argon

Details on Gemini 4 Argon are on the Google Gemini page. Further open model families for self-hosting are listed in the AI models overview, for example Meta Llama or Qwen.

Use cases

GDPR-compliant enterprise AI

• Sensitive data stays in your infrastructure
• No data transfer to external services
• Full control over logging and audit

Specialized applications

• RAG systems: Gemma 4 for the answer, EmbeddingGemma 2 for retrieval, both local
• Multimodal search: search image, video and audio archives with a text query
• Code assistants: internal developer tools and code search
• Customer service: chatbots without data sharing
• Multilingual: TranslateGemma for international teams
• Workflow automation: FunctionGemma for API integration

Edge and mobile

• Gemma 4 E2B/E4B: edge models with thinking mode and audio input
• EmbeddingGemma 2: on-device search with under 600 MB of RAM per Google
• Offline-capable: no internet connection required
• Low latency: local processing

EU availability and GDPR

With open-weights models, the place of operation matters, not the vendor. Gemma can be run entirely in the EU, but the options differ considerably:

Gemma 4 is also available via Hugging Face, Kaggle, Ollama, LM Studio and Docker. The model page of the managed API lists Gemma 4 26B A4B with launch stage “Experimental” (since 3 April 2026), a 262,144-token context and up to 128,000 output tokens. EmbeddingGemma 2 has been announced for the Model Garden; until then it runs via the self-deployment options above. For DiffusionGemma, Google links from gemma.ai directly to the Model Garden.

Integration with CompanyGPT

Gemma models can be integrated into CompanyGPT as a self-hosted option, for enterprises that want to combine Google research with full data control. For search and RAG scenarios, EmbeddingGemma 2 works as a local embedding model, and TranslateGemma suits multilingual environments.

Our Recommendation

Gemma 4 31B is the first choice if you want to run an open model with strong reasoning and code quality yourself: MMLU Pro 85.2%, LiveCodeBench v6 80.0% and thinking mode, under Apache 2.0. For retrieval and search, add EmbeddingGemma 2, especially when images, video, audio or code need to be searchable alongside text.

Depending on the scenario, we recommend:

• Gemma 4 12B Unified for local deployments on laptops or small servers, with audio and 256K context
• Gemma 4 26B A4B for fast inference thanks to 3.8B active parameters, provided there is enough memory for the full model
• Gemma 4 E4B/E2B for edge applications and resource-constrained environments
• DiffusionGemma only for experiments with fast local generation, such as in-line editing; for production quality use Gemma 4
• EmbeddingGemma 2 for local semantic search, on-device RAG and code indexing; in the 270M configuration for text-only workloads
• TranslateGemma for multilingual enterprises without an external translation API
• FunctionGemma as a foundation for your own function-calling models

We currently do not recommend the managed Gemma API on the Agent Platform for personal data from the EU, as it is experimental and processes data in the US. We support you in selecting, deploying and fine-tuning Gemma models in your infrastructure.