Gemma 4 models top Qwen in local setups

Gemma 4's architecture shift marks a major reliability breakthrough for open-weights models operating at the "small" and "medium" scale.

• –Gemma 4 E4B leverages Per-Layer Embeddings (PLE) to deliver the representational depth required for flawless semantic routing.
• –The 26B MoE variant provides reasoning quality competitive with 70B+ models while maintaining the inference speed of a 4B model.
• –Improved "thinking" token efficiency directly addresses the infinite-loop and repetition issues common in competing reasoning models.
• –Native support for agentic workflows and structured output makes this family the new benchmark for local tool-calling pipelines.