【衝撃】14インチMacBook Proで80BのAIが毎秒75トークンを出す。M5 Max MacBook Proの異次元パフォーマンスを検証

First impressions of the 14-inch MacBook Pro (128GB/2TB) with M5 Max. What surprised me most this time wasn't the creative performance, but the local LLM inference speed. Last year's M3 Ultra Mac Studio (192GB) could fit models, but its speed wasn't practical. However, the M5 Max, which has less memory capacity, has achieved usable speeds thanks to its Neural Accelerator and revamped architecture. There's also a live demo of Vibe Coding using LM Studio. ▼ Main Contents 00:00 Intro: From the Disappointment of the M3 Ultra to the Impact of the M5 Max 01:30 Shocking Actual Measurements: Qwen3 Coder Next 80B → 75 Tokens per Second 03:30 Fusion Architecture Explained + The Meaning of a Geekbench Score of 29,430 06:00 Neural Accelerator: The Key to 3.6x Speedup in Prompt Processing 08:00 128GB Unified Memory vs. RTX 5090 (32GB VRAM) 10:00 ★ LIVE DEMO: LM Studio × Qwen3 Coder Next (Vibe Coding Demonstration) 17:00 Three Benefits of Local Vibe Coding 18:30 Creative App Performance (DaVinci / Premiere / Blender) 20:30 The Benefits of SSD Speeds Over 7,000MB/s 21:30 Thermal Limitations of a 14-inch Chassis and the Reality of LLM Operation 23:30 Liquid Glass + 120Hz + Studio Display XDR 26:00 Is ¥848,800 justified? Comparison with Alternatives 28:00 Expectations for the Desktop M5 (Mac Studio x M5 Ultra) ▼ Test Machine Configuration M5 Max (18-core CPU / 40-core GPU) 128GB Unified Memory (614GB/s) 2TB SSD (Over 7,000MB/s) Nano-texture Display Price: ¥848,800 (tax included) ▼ LLM Used for Testing ・Qwen3 Coder Next 80B (6-bit/8-bit quantization) ・Hermes 4 70B ・Execution Environment: LM Studio / MLX Framework #M5Max #MacBookPro #AppleSilicon #LocalLLM #VibeCoding #AIWorkstation --- Masakazu Honda's Weekly Tech Talk A journalist with over 30 years of experience in the tech industry explains the "essence" of the latest technology every week.