During the pretraining stage, the team used six capability probes -- scientific reasoning, code, factuality, multilingual, vision, and reasoning -- to ensure well-rounded model development, with multilingual ability showing the most significant improvement due to richer corpora and balanced training. On standard benchmarks, the Gemma 3 27B-parameter version delivered outstanding performance: achieving 89.0% on MATH, approaching Gemini 1.5 Pro; and attaining leading or comparable results to larger closed-source models on MMLU-Pro, LiveCodeBench, Bird-SQL, GPQA Diamond, FACTS Grounding, and MMMU.
It supports at least 128K-token contexts, achieving 66.0% accuracy on the RULER 128K benchmark, demonstrating strong long-document processing capability. With the integrated SigLIP vision encoder, its performance on document understanding tasks (such as DocVQA, InfoVQA, TextVQA) even surpasses the larger PaliGemma 2 model, and the Pan & Scan image processing method improves robustness when handling images of varying aspect ratios. In terms of efficiency, the 27B model can perform efficient inference on a single H100 GPU, with computational requirements only about 1/10 of comparable models, and it uses a 5:1 interleaved local/global attention layer architecture to optimize memory usage under long contexts. On the safety front, its memorization rate is significantly lower than previous generations, harmful knowledge levels in CBRN knowledge assessments are shown to be extremely low, and scans revealed no personal information leakage.