Gemma 4 hits production-ready open weights

Gemma 4's performance signals that open-weight models are finally closing the gap with proprietary giants for medium-complexity production tasks.

• –Vision performance standout, correctly identifying single points of failure in architecture diagrams and extracting dense data from complex charts.
• –The 26B MoE model achieves near-parity with the 31B dense version while using only 3.8B active parameters per token, offering incredible efficiency for edge deployment.
• –Multi-file code generation reveals a slight "knowledge lag," with the model occasionally using deprecated FastAPI handlers and mixed Pydantic syntax.
• –Thinking Mode support in the 31B model allows for deeper reasoning on complex logic puzzles that previously required proprietary APIs.
• –Results independently verified by Gemini 3.1 Pro and Claude Opus 4.6, adding significant credibility to these "non-benchmark" real-world evals.