Readers testing Gemma 4 locally report that while the model delivers exceptional inference speed and impressive one-shot coding performance rivaling top-tier proprietary AI, it struggles significantly with agentic tasks, tool use, and large context windows. Although early benchmarks were skewed by tokenizer and tool invocation bugs, corrected tests show the dense 31B variant actually outperforms the faster MoE model in overall quality. Ultimately, users praise its capabilities for specific local coding tasks on Apple Silicon but note that its limitations in iterative reasoning and context scaling prevent it from being a flawless autonomous coding assistant.