Introducing Gemma 3

Source: DeepMind Date: 2025-03-12 URL: https://deepmind.google/blog/introducing-gemma-3/

Summary

Google released Gemma 3 in 1B, 4B, 12B, and 27B parameter sizes with 128K context window, multimodal vision, function calling, and 140-language multilingual support. The 27B variant claims Chatbot Arena Elo 1338 — outperforming Llama 3-405B, DeepSeek-V3, and o3-mini in preliminary human preference evals — while requiring only a single GPU (vs. competitors at up to 32 GPUs). Includes ShieldGemma 2 for image content moderation.

Implications

The open-weights thread. Gemma 3 27B outperforming 405B-class models on Arena Elo on a single GPU is the efficiency claim that matters for the open-weights market. If real, it shifts the deployment calculus for anyone running local or small-cloud inference — you don’t need the big model anymore. That competes directly with Llama 3.3 70B and Mistral Large as the “good enough at reasonable cost” tier.

Multimodal open-weights are rarer. Vision + function calling + 128K context at open weights is a meaningful bundle. Most open-weight competitors (Llama 3.1 at time of release) didn’t include native vision. Gemma 3 closes that gap and makes Google’s open-weight offering more credible for production agentic use cases.

ShieldGemma 2 as safety tooling. Shipping a content moderation model alongside the main release is Google’s attempt to make safety tooling part of the open-weights distribution. If ShieldGemma 2 is good enough, it becomes the default safety layer for Gemma-based applications — extending Google’s influence into the open-source safety layer.

Watch:

• Whether Gemma 3 27B Elo holds on independent community evals (not Google’s preliminary numbers)
• Llama 4 and Mistral response — the open-weights competitive timeline accelerates with each Gemma release
• ShieldGemma 2 adoption for content moderation in production Gemma deployments