Gemma
Free Last verified: September 2026Google's open-weight model family, with the multimodal Gemma 4 generation released under Apache 2.0
What is Gemma?
Gemma is Google DeepMind's family of open-weight language models designed to give developers and researchers access to the same research lineage behind the flagship Gemini models, packaged as models you can download, fine-tune, and deploy however you want. Unlike Gemini itself (which remains closed-weight and API-only), Gemma 4 is released under the Apache 2.0 license, one of the most permissive open-source licenses in existence, which means you can use it commercially, modify it, redistribute it, and self-host it without paying Google or asking permission. Note that this is a Gemma 4 change: Gemma 1 through 3 and the specialist variants are still governed by Google's custom Gemma Terms of Use, and Google's Prohibited Use Policy applies to every generation. The Gemma 4 generation, released in April 2026, is multimodal and comes in five sizes: E2B and E4B for phones and edge devices, a 12B, a 26B mixture-of-experts model with roughly 3.8B active parameters, and a 31B dense flagship that competes with much larger open models on benchmarks. Every Gemma 4 model natively handles images and video alongside text at variable resolutions, while native audio input is limited to the E2B and E4B checkpoints. Context runs to 128K on the edge models and up to 256K on the larger ones, across 140+ languages. Gemma 3 lives on as a mature, stable option for production workloads that prioritize reliability over frontier performance. Where Gemma really differentiates is in its optimization for on-device and edge deployment: Google ships official Gemma packages for MediaPipe, Keras, JAX, PyTorch, and Hugging Face Transformers, plus a deep bench of specialist variants. The current lineup includes FunctionGemma for tool calling, EmbeddingGemma for on-device embeddings, ShieldGemma 2 for safety classification, MedGemma 1.5 for medical text and imaging, TranslateGemma for multilingual translation, VaultGemma for differentially private training, T5Gemma 2, and DiffusionGemma; the older CodeGemma and RecurrentGemma checkpoints remain downloadable but are legacy Gemma 1 era releases. If you want a free, safety-conscious open model with Google-grade pretraining data and a genuinely permissive license, Gemma is the obvious pick alongside Llama and Mistral.
Gemma demo video
Watch Google for Developers's official demo to see Gemma in action before reading our full review.
Official video by Google for Developers via YouTube, embedded for reference. ToolChase does not host or claim this video.
⚡ Quick Verdict
Teams that want an Apache-licensed open model with Google-grade training and strong on-device deployment options
Users who want the absolute largest open model or the biggest community ecosystem
Free to download · Vertex AI API from $0.14 input / $0.40 output per million tokens
Yes, all weights are free to download; Apache 2.0 covers Gemma 4, earlier generations stay under the Gemma Terms of Use
Apache 2.0 licensing plus Google-quality pretraining, the most commercially clean open model family
Smaller ecosystem and flagship size than Llama
Bottom line: Gemma scores 4.3/5, the top pick when you need an Apache-licensed open model or efficient on-device deployment. Go Gemma 4 31B-IT for server, 26B MoE for a consumer GPU, and E2B/E4B for mobile.
Pricing
Weights, Free: Every Gemma generation is free to download from Hugging Face or Kaggle, and Google's terms confirm that "You may reproduce or Distribute copies of Gemma or Model Derivatives" and permit commercial use, with no fee anywhere. The license is not uniform across the family, though: Google's terms page says "For Gemma 4 terms, see the Gemma 4 license" (Apache 2.0), while "the terms below apply to Gemma models listed in the Appendix", so Gemma 1, 2, 3 and the specialist variants stay under the custom Gemma Terms of Use with the Prohibited Use Policy attached. Note also that Google's Gemma page has moved: ai.google.dev/gemma now returns a 301 redirect to deepmind.google/models/gemma, which describes Gemma as "Our most capable open models" but carries no license or pricing detail.
Google Cloud Vertex AI (if using managed API): Gemma 4 31B-IT from roughly $0.14 per million input tokens and $0.40 per million output tokens. Smaller Gemma models are proportionally cheaper.
Third-party inference: Groq, Together AI, and Fireworks serve Gemma models at competitive per-token rates, typically $0.05-$0.50 per million tokens depending on size. Self-hosting is free beyond hardware costs, and the Gemma 4 E2B and E4B checkpoints run on consumer laptops.
Verification note, 2 August 2026: the weights being free under Apache 2.0 is confirmed on Google's own Gemma terms page. The per-token figures above are carried over from our May 2026 check and we could not re-confirm them at the source: Google's Vertex AI generative AI pricing page does not list Gemma among its per-token models, which suggests Gemma on Vertex is billed as self-deployed compute rather than per token. Third-party host rates also move constantly. Treat every per-token number on this page as indicative and price your specific provider before committing.
Key Features
- Apache 2.0 license on Gemma 4, commercial use permitted subject to Google's Prohibited Use Policy
- Gemma 4 multimodal family, image and video on every size, native audio on E2B and E4B
- Five sizes from E2B (mobile) up to 31B dense (server-class), including a 26B MoE
- 128K context on edge models, up to 256K on the larger ones, 140+ languages
- FunctionGemma for tool calling and EmbeddingGemma for on-device embeddings
- Official Keras, JAX, PyTorch, and Transformers support
- MediaPipe integration for on-device mobile deployment
- Strong safety tuning and Responsible AI Toolkit
Pros & Cons
Pros
- Apache 2.0 license is genuinely permissive, easier than Llama
- Excellent on-device story with MediaPipe and 1B/2B sizes
- Backed by Google DeepMind research and pretraining quality
- Multimodal Gemma 4 competes with much larger closed models
Cons
- Ecosystem smaller than Llama, fewer fine-tunes and tools
- Flagship Gemma 4 31B is smaller than Llama 3.3 70B
- Less community-driven support than Mistral or Qwen
FAQ
Is Gemma really Apache 2.0?
Gemma 4 is, but the rest of the family is not, so check the generation before you rely on it. Google moved to a standard Apache 2.0 license for Gemma 4 in April 2026, with no monthly-active-user threshold and no attribution requirement beyond normal Apache notices. Gemma 1, 2, and 3, along with specialist variants such as PaliGemma, ShieldGemma, and FunctionGemma, are still distributed under the custom Gemma Terms of Use. One caveat applies either way: Google's Prohibited Use Policy is incorporated by reference into the Gemma terms, so there are restrictions on what you may build, and you must pass them down to anyone you redistribute to. On Gemma 4 specifically there is no hidden scale clause to worry about.
How does Gemma 4 compare to Llama 4?
Llama 4 Maverick and Scout are larger (17B active / 109B-400B total) and score higher on most reasoning benchmarks. Gemma 4 31B is smaller but punches above its weight on multimodal and coding benchmarks, and its Apache 2.0 license is more permissive than Llama's community license. For most production workloads, Llama still wins on raw capability, but Gemma is often preferred for on-device and strict-licensing deployments.
Can I run Gemma on my laptop?
Absolutely, and Gemma 4 is explicitly built for it. The E2B and E4B edge checkpoints run comfortably on a modern MacBook or any machine with 8-16GB of RAM via Ollama, LM Studio, or llama.cpp. The 12B works well with a consumer GPU like an RTX 4070 or a Mac with 24GB of unified memory, and the 26B mixture-of-experts model is designed to fit consumer GPUs because only about 3.8B parameters are active per token. The 31B dense flagship is the workstation and server option and wants an A100 or H100, or two consumer cards with quantization.
What is CodeGemma?
CodeGemma is a specialized Gemma variant fine-tuned for programming tasks, code completion, infilling, and instruction following for coding. It comes in 2B and 7B sizes and is optimized for low-latency code assistance scenarios like IDE plugins and inline completion. CodeGemma competes with StarCoder2 and DeepSeek Coder in the open-weight coding model category.
Where can I use Gemma via API?
Google Cloud Vertex AI hosts Gemma with official Google SLAs. Third-party providers including Groq, Together AI, Fireworks, and OpenRouter all serve Gemma at pay-per-token rates. You can also run it locally with Ollama or Hugging Face Transformers for free.
Does Gemma support function calling?
Yes, and the story is better than it was. Gemma 4 supports tool use and structured JSON output directly, and Google now ships FunctionGemma, a variant purpose-built for function calling, so you no longer have to lean on prompt-level JSON coercion for agent workflows. The surrounding tooling is still thinner than Llama's or OpenAI's, so expect to do more integration work, but tool use is no longer an afterthought in the Gemma line.
Is Gemma safe for sensitive data?
When self-hosted, Gemma keeps all data on your own infrastructure, ideal for healthcare, finance, and government workloads where closed APIs are a compliance barrier. Google also publishes a Responsible AI Toolkit with Gemma, including safety classifiers, prompt filters, and red-teaming guidance.
📋 Good to know
Download weights from Hugging Face or Kaggle, run with Ollama, Keras, Transformers, or MediaPipe for mobile.
Self-hosting keeps all data local. Google Vertex AI usage subject to standard GCP data policies.
Move from Gemma 4 E2B/E4B to the 12B, 26B MoE, or 31B dense when you need reasoning capability.
Low, Ollama and Hugging Face Transformers work out of the box with one-line installs.
