MediaTek's 2nm Dimensity 9600 Pro runs 30B AI models on your phone
First 2nm flagship silicon promises 40% faster tokens and an on-device agent engine — plus a Claude Code update that locks Bash to per-command domains.
Copy markdown
30B models, on the phone in your pocket
The Dimensity 9600 Pro's dual NPU 1090 runs mixture-of-experts models up to 30B parameters entirely on-device, with 51% higher LLM prefill than last gen. That's a mobile target finally big enough for a real local agent — no API round-trip.
The inference math: 40% faster tokens, 55% more per watt
MediaTek claims 40% faster token generation and 55% higher tokens-per-watt, paired with first-ever LPDDR6 memory (33% more bandwidth) and UFS 5.0 storage. Local inference stops being a battery-drain party trick and starts being usable all day.
An on-device agent engine, not just a chatbot
A new Agentic AI Engine is pitched for on-device data analysis, proactive assistance, and preference learning — the phone-side half of the agent stack. If you build mobile agents, the silicon to keep context and tools local now exists.
The catch: MoE, not dense — and no phone until Q4
That 30B ceiling is sparse MoE, where only a few billion params fire per token, not a dense 30B. And you can't build against it yet: the first Dimensity 9600 Pro phones are due 'this quarter,' with no dev kit announced.
Elsewhere: Claude Code 2.1.271 locks Bash to per-command domains
The newest Claude Code reviews Bash, PowerShell and Monitor commands with only their required hosts in sandboxed auto mode, adds fast mode to remote sessions, and renders published Markdown as styled document pages. Tighter blast radius for agentic shell work.