🧠Gemma 3 Outperforms Larger Models
Google DeepMind released the Gemma 3 family on May 6, 2026, with the 27-billion parameter variant delivering benchmark scores that surpass Meta's Llama 3 70B, a model nearly three times its size. Gemma 3 27B scores 75.8 on MMLU versus Llama 3 70B's 73.1, 82.3 on HumanEval versus 79.2, and 64.5 on MMMU (multimodal understanding) versus 58.3. The model achieves this through knowledge distillation from a 300B Gemini-family teacher model and an architecture that uses grouped-query attention with sliding windows to reduce memory footprint while maintaining quality.
The Gemma 3 family includes variants at 1B, 4B, 12B, and 27B parameters, all sharing the same tokenizer and architecture. The 1B and 4B variants are designed for on-device deployment and ship pre-optimized for MediaTek Dimensity 9400 SoCs, enabling offline AI features on Android devices. Google claims the 1B model can run at 60 tokens per second on a Pixel 11 with 2GB of RAM using 4-bit quantization.
đź“‹ Vision, Audio, and Multilingual Support
All Gemma 3 models embed a SigLIP vision encoder that enables image understanding across resolutions from 224px to 896px. Users can input photos, screenshots, and documents for visual question answering, OCR, and chart interpretation without a separate vision model. Google demonstrated Gemma 3 27B reading a handwritten medical prescription, extracting structured data, and cross-referencing drug interactions—all locally on a workstation with 32GB VRAM.
The models support over 140 languages in a single checkpoint with a unified 256K vocabulary SentencePiece tokenizer, avoiding the common pattern of separate models for different language groupings. Multilingual benchmarks show Gemma 3 27B achieving MGSM scores of 82.1 (averaged across 12 languages), compared to Llama 4 70B's 81.7. The 128K context window enables processing of long documents like legal contracts or research papers in their entirety.
🛡️ Safety Tools and Intellectual Property Indemnification
Google includes ShieldGemma 2, a 4B-parameter safety classifier trained to detect sexually explicit content, hate speech, harassment, and dangerous content (weapons, CBRN instructions), at no additional cost. ShieldGemma 2 can be deployed as a content filter before user-facing outputs, and Google reports a 97.2% recall rate on policy-violating content in adversarial testing.
In a significant legal move, Google extended its intellectual property indemnification to commercial users of Gemma 3, covering copyright claims related to the model's training data. This follows similar moves by Microsoft for Copilot and Adobe for Firefly, and addresses a key barrier for enterprise adoption of open models—the unresolved legal status of training data. Google specified the indemnity covers Gemma 3 models used under the standard Gemma terms of service without modification.