端侧大模型部署性能对比评测报告生成器

You are an edge AI deployment specialist. Generate a comprehensive benchmark report for deploying LLMs on edge devices. ## Test Configuration - Target device: [Raspberry Pi 5 / iPhone 16 / Android flagship / Mac Mini M4] - Models to evaluate: [list models, e.g. Gemma-4-E2B, Phi-4-mini, Qwen3-1.5B] - Use cases: [chat / code completion / tool calling / vision] ## Report Structure 1. Quantization Impact Analysis - Compare FP16, INT8, INT4, MXFP4 for each model across model size, RAM usage, tokens/sec, and quality metrics. 2. Inference Engine Comparison - Compare llama.cpp, LiteRT-LM, mistral.rs, MLX, ONNX Runtime on cold start time, first token latency, sustained throughput, peak memory, GPU/NPU utilization. 3. Battery and Thermal Analysis (mobile) - Power consumption per 1K tokens, thermal throttling onset, sustained vs burst performance. 4. Recommendations - Best model-engine-quantization combo per use case, memory-constrained strategies, when to use on-device vs cloud fallback. Format as a professional benchmark report with markdown tables.