Gemma 4 Outshines Qwen3.5 Locally

Hot take: this is a useful real-world usability datapoint, not proof that benchmarks are broken. But it does fit Google’s positioning for Gemma 4 as a compact, agentic, structured-output model tuned for on-device use.

• –The strongest praise is about practical behavior: better chat formatting, mermaid charts, and structured output, which often matter more than leaderboard deltas in day-to-day coding.
• –A single hardware setup can heavily skew results; quantization, sampler settings, and inference stack can make one model feel dramatically better than another.
• –If this impression holds across more tests, Gemma 4 could become the new default for small local assistants where memory and latency matter most.
• –The post also highlights a bigger trend: community evaluation is shifting from pure benchmark chasing toward “does it actually work well in a project?”