联发科天玑 9600 Pro 双 NPU 可跑 30B 本地模型
联发科发布 2nm 天玑 9600 Pro，称 NPU 1090 支持端侧 30B 参数模型推理，另有常驻低功耗 NPU 跑 AI 代理，但尚无真机实测数据。
Your next phone could run a 30B AI model on its own (No Cloud). 🎯So laptops and home PCs are not the only domain for Local AI. MediaTek launched the 2nm Dimensity 9600 Pro says its new NPU 1090 can run on-device AI applications with models up to ... 🧠 30 BILLION parameters This chip looks specifically designed around Local AI ... ⚡ 2× INT4 compute vs previous generation 🚀 +51% LLM prefill performance 🔋 +55% tokens generated per watt 🧠 Hardware KV-cache compression 🧩 On-device MoE inference 💾 LPDDR6 memory 💽 UFS 5.0 storage There are actually 2 x NPUs. 🔥 NPU 1090 → heavy generative AI / LLM inference 💤 Super Efficient NPU 2.0 → low-power, always-on AI agents MediaTek wants your phone's AI to do more than answer prompts. It wants it running continuously, analyzing on-device data, learning preferences, and acting as a local personal agent. Laptop Local AI → Mini PC Local AI → Phone Local AI. A 30B-class model in your pocket with no API call may no longer sound ridiculous. 👀 ⚠️ MediaTek has not shown that you can easily download any 30B GGUF from Hugging Face and run it . We still need to see real phones + real models + real tok/s. 🔗 Announcement in ALT