MediaTek's 2nm chip puts a 30B MoE model on your phone
The edge is the story today: new silicon built for agents, an OS safety suite that runs its models on the device, and a fresh open model betting on spatial reasoning — all landing within hours of each other.
MediaTek launched the Dimensity 9600 Pro, a TSMC 2nm flagship whose headline spec is an agent, not a benchmark. The chip pairs a dual-NPU design with an "Agentic AI Engine" and a second-generation efficiency NPU fused to a low-power sensor hub — a 40% cut in always-on AI power that is what lets a phone assistant perceive and act continuously instead of waking on a tap. The performance NPU (1090) delivers 51% higher LLM prefill throughput and up to 55% more generated tokens per watt, with INT4 compute doubled, hybrid-attention support, and new compression that lets models up to 30B parameters (MoE) run entirely on-device. The rest of the silicon is built around it: a 2+3+3 all-big-core CPU topping out at 4.55 GHz with 17% higher single-core and 61% lower multi-core power at last-gen peak, a 34.5 MB cache, and the first LPDDR6 memory and UFS 5.0 storage in a phone chip. Android Police notes it beats Qualcomm to the 2nm finish line; MediaTek used the launch to argue the whole market is becoming a token market, citing an estimate of 120 quadrillion tokens consumed globally per month by 2030 — 24x current levels, per the company's own stage figures. The strategic read: after the news this morning that China's new five-year plan puts the AI model on the device, the silicon vendor with the biggest mid-tier phone footprint is betting that on-device agents, not cloud chatbots, are the upgrade cycle.
Apple's iOS 27 child safety suite is now live — with gore detection, on-device models, and a contradiction. Communication Safety, previously limited to nudity detection, now blurs and warns on gore and violent content in images and videos shared with children under 18, on by default, with all classification running on the device so Apple says it receives no indication when something is detected. Parents also get a new slider to partially reveal a blurred image without viewing it in full, Safari's "Ask to Browse" routes a child's visit to any new website through parental approval, and Apple is working with the American Academy of Pediatrics on a Family Media Plan guide. The wrinkle: Techmeme's pick says the protections cover live FaceTime calls, but Apple's own support document, published the same day as the announcement, says gore and violence detection works "only in FaceTime video messages, not live calls," and isn't on Vision Pro at all. The Mac Observer laid out all four conflicting Apple pages. Until Apple says which description matches the shipped software, parents should assume the narrower one and rely on contact limits for live calls.
Zidong Taichu open-sourced a 9B multimodal model built around space, not captions. The CAS Institute of Automation team behind the Zidong Taichu (紫东太初) series released ZDTaichu5.0-9B, which combines a Qwen3.5-9B language backbone with NVIDIA's C-RADIOv4-H vision encoder and takes text, images, and video at any resolution. Its claim is the narrowest useful one in the current small-model wave: first-tier general visual understanding, plus the strongest spatial and embodied reasoning of any general-purpose VLM under 10B parameters, with agentic tool use included. Weights are on Hugging Face and ModelScope under the NVIDIA Open Model License, with an adapted vLLM fork and a recurrent-reasoning component in the repo. Spatial reasoning is the capability cloud-scale models keep failing at and the one robots actually need — this is the same bet the MediaTek chip above is made to serve, from the model side.
What to watch: whether an always-on phone agent survives contact with real battery life once the first 9600 Pro devices ship.
Is on-device AI the escape hatch from API pricing, or just a worse version of the same model at a fifth of the scale? Tell us in the comments.
Sources: ZDTaichu5.0-9B (GitHub)
