Gemma

Gemma is a family of open-weight large language models developed by Google DeepMind, first released in February 2024. Built on the same research and technology as Google's flagship Gemini models, Gemma offers a lightweight, accessible alternative that developers and researchers can download, customize, and deploy on their own hardware.

By mid-2026, over 150 million downloads of Gemma models had been recorded, with more than 70,000 community variants available on Hugging Face. The latest generation, Gemma 4, was released on April 2, 2026 under the permissive Apache 2.0 license — a significant shift from the source-available license used by earlier versions.

Google debuted Gemma in February 2024 as a collection of source-available LLMs serving as a lightweight version of Gemini. The initial release came in two sizes: 2 billion and 7 billion parameters. Multiple publications viewed this as a response to competitors like Meta releasing open-source AI models, marking a shift from Google's longstanding practice of keeping its AI source code private.

Gemma 2 was released on June 27, 2024, offering 2B, 9B, and 27B parameter variants with improvements in efficiency and performance. Gemma 3 arrived on March 12, 2025, introducing multimodal capabilities (text and image input), support for over 140 languages, and a 128K context window across 1B, 4B, 12B, and 27B parameter sizes.

Gemma 4 was released on April 2, 2026 under the Apache 2.0 license. This release introduced a 31B Dense model that reached third place on Arena's text leaderboard, along with a 26B Mixture of Experts (MoE) model, 12B model, and smaller edge-optimized variants. On June 3, 2026, Gemma 4 12B was released with a unified multimodal architecture that processes images and audio without dedicated encoders.

Gemma models use a decoder-only transformer architecture similar to other modern LLMs. Key architectural features include:

• Grouped-Query Attention (GQA) — improves inference efficiency by reducing KV-cache size
• SigLIP Vision Encoder — used for multimodal image understanding (Gemma 3+)
• 128K-256K Context Window — long context support across most model sizes
• Quantization-Aware Training (QAT) — official quantized variants with minimal accuracy loss
• Multi-Head and Multi-Query Attention — different attention mechanisms across model sizes

Beyond the core Gemma models, Google has developed several specialized variants for specific use cases:

Variant	Purpose
CodeGemma	Code completion and generation (2B, 7B)
PaliGemma	Vision-language model combining SigLIP with Gemma
MedGemma	Medical analysis and image understanding
ShieldGemma 2	Content safety — identifies violent, dangerous, and explicit images
TranslateGemma	Machine translation (4B, 12B, 27B)
RecurrentGemma	Griffin-based architecture (non-transformer) with unlimited context
DolphinGemma	Designed for dolphin communication analysis (research only)

Gemma 4 represented a major leap forward. Available in multiple sizes:

• 31B Dense — the flagship model, ranked #3 on Arena's text leaderboard
• 26B A4B MoE — Mixture of Experts architecture, ranked #6
• 12B — unified multimodal (released June 2026, processes images/audio natively)
• E4B (Effective 4B) — optimized for edge devices
• E2B (Effective 2B) — ultra-compact for mobile/on-device deployment

All models support multimodal input (vision, with audio on E2B/E4B). Context windows range from 128K on edge models to 256K on larger variants. All are released under the Apache 2.0 license.

Licensing evolved significantly across Gemma generations. Gemma 1 through 3 used a source-available license (Gemma Terms of Use) that permitted most use cases but restricted certain commercial applications and required registration. Gemma 4 shifted to the fully open-source Apache 2.0 license, allowing unrestricted use, modification, and distribution — including for commercial purposes. This was widely seen as Google's response to Meta's Llama and Mistral's increasingly permissive licensing.

Aspect	Gemma	Llama	Mistral
Developer	Google DeepMind	Meta	Mistral AI
Latest Version	Gemma 4 (Apr 2026)	Llama 4 (2026)	Mistral 4 (2026)
Max Size	31B Dense, 26B MoE	405B Dense	123B MoE
License	Apache 2.0 (Gemma 4)	Custom	Apache 2.0 / Custom
Multimodal	Vision + Audio (Gemma 4)	Vision (Llama 4)	Vision (Pixtral)
Context	128K-256K	128K	32K-256K
Ecosystem	Keras, JAX, Hugging Face	PyTorch, Hugging Face	Ollama, Hugging Face

• On-device AI — small Gemma variants (2B, 4B) are ideal for smartphones, laptops, and edge devices
• Medical AI — MedGemma enables privacy-preserving medical image analysis on local hardware
• Code Generation — CodeGemma provides specialized code completion and generation
• Multilingual NLP — Gemma 3 supports 140 languages for global applications
• Vision-Language — PaliGemma combines text and image understanding for document processing
• Content Moderation — ShieldGemma 2 filters violent and explicit image content

Gemma is Google's strategic bet on open-weight AI — a counterbalance to its closed Gemini models. The shift to Apache 2.0 with Gemma 4 makes it one of the most permissively licensed capable models available. Its strength lies in offering high-quality small-to-medium models that punch above their weight, particularly in edge and on-device scenarios. While it doesn't compete with the largest frontier models like GPT-5 or Claude Opus at the high end, Gemma's ecosystem of specialized variants (CodeGemma, PaliGemma, MedGemma) and its strong developer tooling make it an excellent choice for builders who want customizable, deployable AI without vendor lock-in.