Gemma

Gemma is Google's open-source model family, built on the same research as Gemini. While Gemini powers Google's consumer products, Gemma is designed for developers who need deployable, self-hostable models that run efficiently on limited hardware — laptops, phones, or small cloud instances.

Model variants (as of August 2026): The current flagship is Gemma 4, available in 12B, 26B, and 31B parameter sizes, plus ultra-compact E2B and E4B variants optimized for maximum compute/memory efficiency on-device. Gemma 3 includes a 270M hyper-efficient model. Specialized variants cover a wide range of domains: DiffusionGemma (fast text diffusion), T5Gemma 2 (encoder-decoder), MedGemma 1.5 (medical imaging/text), TranslateGemma (55-language translation), FunctionGemma (edge function-calling), ShieldGemma 2 (content safety), VaultGemma (differential privacy), and EmbeddingGemma (on-device embeddings).

Who it's for:

• Developers building AI features who want to avoid API costs at scale
• Researchers needing a capable open model for experimentation
• Companies with data privacy requirements that preclude sending data to external APIs
• Edge computing applications where latency matters

Running Gemma: Models are available on Hugging Face, Google's Kaggle, and Vertex AI. Tools like Ollama, LM Studio, and llama.cpp support Gemma locally on consumer hardware.

Performance: Gemma 4's 31B model targets frontier-level capability in a mobile-first footprint, per Google's own positioning — a significant jump from the Gemma 2 generation.

License: Gemma uses the Gemma Terms of Use (not Apache 2.0). Commercial use is permitted but with restrictions around certain large-scale deployments.

Alternatives: Meta's Llama 4 (Scout/Maverick) is the most direct competitor with a more permissive license. Microsoft's Phi family is similarly size-optimized. Mistral's open models offer strong performance in the same tier.