Gemma Glossary

Google DeepMind's family of open-weight models, designed for deployment and adaptation outside Google's hosted services.

Gemma is the open-weight model family released by Google DeepMind alongside the proprietary Gemini line. The first Gemma models appeared in February 2024. Gemma 4, released in March 2026, included dense and mixture-of-experts variants, multimodal input, and context windows of up to 256,000 tokens.

For this Dictionary, Gemma matters because a laboratory with a major closed-model business also releases weights that operators can run and adapt elsewhere. That makes the family useful for practical Sovereign Compute experiments without pretending that “open weights” settles every question about training data, licensing, or reproducibility.

In the operator’s May 2026 tests on Apple Silicon, Gemma 4 was fast on contained tasks, while Qwen was more reliable on some large-context, multi-file work. That is a dated local observation, not a universal model ranking: quantisation, runtime, prompt construction, and hardware all affect the result.

Sources

• Google, Gemma releases.
• Google, Gemma 4 model overview.

See also

• Google DeepMind
• Gemini — the proprietary sibling line
• Sovereign Compute
• Hermes, Qwen, Llama — peers in the local-compute tier

Return to Dictionary All Entries (A–Z) For Students Other Writing Capstone 2.0