Gemma 3 is well documented by Google DeepMind as an open-weight family with 1B, 4B, 12B, and 27B variants, 128K context, broad multilingual coverage, text/image input with text output, function calling, quantized releases, and local or cloud deployment paths. This gives it strong technical breadth and developer usability for an open model family, especially where deployment control matters. Coding evidence is mixed and materially below frontier coding agents. Google’s model card and technical report provide LiveCodeBench results, with the strongest cited Gemma 3 IT result reaching 39.0 in one table. Aider lists gemma-3-27b-it at 4.9% on its polyglot benchmark, DeepSWE has no Gemma 3 entry, and supplied SWE-bench evidence does not show a Gemma 3 Verified score. Preference evidence is also moderate: Arena ranks gemma-3-27b-it relatively low overall, while Arena-Hard shows some usable general and creative-writing signal. Cost and availability are strong because the model is open-weight, appears in multiple runtime/provider contexts, and Artificial Analysis lists Google AI Studio pricing at $0.00 for Gemma 3 27B. Pricing clarity is not complete: supplied Google Cloud evidence covers supervised fine-tuning prices, while runtime costs depend on provider or self-hosting. Overall evidence quality is good for capabilities and deployment, but weaker for independent benchmark coverage and exact licensing/pricing details.
