Google's Gemma 4 family arrives as open-weight models that can see, hear, and — at least in a limited sense — think.
The Gemma 4 suite spans 2.3B to 31B parameters and comes in both dense and Mixture-of-Experts architectures. All model sizes get improved vision and audio encoders; the 12B variant goes further with an encoder-free design that ingests raw audio and image patches directly. Google also adds a "thinking mode" across the family, which generates explicit reasoning traces before producing a final answer — a pattern that larger proprietary models have been using to boost performance on complex tasks. The technical report claims gains on STEM, multimodal, and long-context benchmarks, with results described as rivaling larger frontier open models on human-rated evaluations.
The architecture choices matter because they push capable multimodal reasoning into a weight range that researchers and developers can actually run locally or fine-tune cheaply. Open-weight models at this scale have historically lagged proprietary ones on vision and audio tasks; if the benchmark claims hold under independent testing, that gap narrows meaningfully. The thinking-mode addition is also notable — it follows a pattern set by models like DeepSeek-R1 and brings structured chain-of-thought reasoning to a family previously without it.
Benchmarks in a self-published technical report are a starting point, not a verdict — independent evaluations will tell a clearer story.