Gemma Glossary

Google DeepMind's open-weights model family, derived from the Gemini lineage. The operator runs Gemma 4 26B and 31B locally on Apple Silicon.

Gemma is the open-weights model family released by Google DeepMind, structured as a sibling to the proprietary Gemini line. The first Gemma generation was released in February 2024; subsequent generations (Gemma 2, Gemma 3, Gemma 4) have shipped at increasing scales and capabilities. As of May 2026, the operator of this Dictionary runs Gemma 4 26B and Gemma 4 31B locally on the M5 Max workstation as part of the practical Sovereign Compute setup — see TOOLS.md notes on hybrid dense+MoE architecture and sliding-window attention.

For this Dictionary, Gemma matters as the open-weights family produced by an actor with frontier-lab resources and full-Gemini-team backing. Where Mistral and Llama are produced by labs that have made the open-weights posture their primary strategic move, Gemma is produced by a lab with closed-weights frontier products that has chosen, deliberately, to release competitive open-weights alongside them. The behaviour is asymmetric to OpenAI (which has released essentially no open weights since GPT-2) and Anthropic (no open weights at all) and somewhat parallel to Meta’s Llama posture.

Strengths of Gemma 4 relative to peer open-weights models (Hermes, Qwen) on local Apple Silicon: fast on contained tasks, punches above weight on moderate-context workloads. Weaknesses: the sliding-window attention pattern limits the model’s behaviour on large-context multi-file tasks, where Qwen 3.6 currently outperforms.

See also

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