Arm's new mobile chip platform is built for on-device AI agents

Arm has redrawn the blueprint that many smartphone chips start from, and it did so with AI agents in mind. On 8 September 2026 the company introduced CSS for Mobile 2, a pre-integrated platform that silicon partners license and build their own phone chips from.[1][2]

Arm's argument is that a phone agent does more than answer questions. In Arm's description, it must keep context, run apps, coordinate models and services, and act for the user, all inside a smartphone's power and heat limits.[1]

The platform bundles several layers. The Arm C2 CPU cluster pairs C2-Ultra, Arm's most powerful mobile CPU, with efficiency-focused C2-Pro cores and two units of SME2, Arm's matrix extension for AI work. The Mali G2-Ultra NX GPU is the first Mali graphics chip with dedicated neural accelerators. System IP, physical implementations and software complete the package.[1]

Chipmakers are not forced to take everything. Arm's product blog says the components can be used on their own or combined with custom and third-party designs. It also says the platform reaches into physical implementation, so partners can weigh performance, power and chip area before they begin integrating a system-on-chip.[2]

Against the previous C1-Ultra, Arm says C2-Ultra delivers up to 1.7x higher AI performance and 15% higher single-thread performance, while using up to 38% less power at the same performance. Its blog adds 15% faster web browsing, 12% faster app launch and a sample agent workflow of speech, retrieval, reasoning, app use and browsing that finished 24% faster.[1][2]

For graphics, Arm claims up to 4x higher performance per watt for neural graphics and up to 14% higher performance on existing games. Every figure is Arm's own. Phone maker vivo says it is bringing Arm Neural Technology to its latest flagship phones built on Arm's newest platform, but the announcement names no model or shipping date.[1][2]

Arm casts the CPU as the orchestration engine. In its account, the processor keeps track of a task, gathers context, decides whether each step runs in a local app, an on-device model, another processor or a cloud service, and coordinates permissions. Arm says doubling the SME2 units gives the CPU more headroom to run AI itself while working alongside accelerators.[2][1]

Software is the other half of the pitch. Arm points developers to its KleidiAI libraries, framework integrations and a new Arm AI Portal listing optimised models with performance and accuracy data, and says more than 22 million developers build on Arm. Its graphics blog counts more than 14 billion Mali GPUs shipped.[2][3]