A new SDK builds local-only assistants for regulated sectors like healthcare and finance, pitching on-device inference as a privacy and compliance win.

The company launched a developer kit for building assistants that run locally, targeting healthcare, finance, and other sectors with strict data-handling rules. Keeping inference on-device is pitched as a way to ease privacy and compliance concerns.

Analysts see spend moving from one-off training to ongoing inference — changing how operators budget and plan hardware.

A research note argues spending is steadily tilting from one-time model training toward the recurring cost of serving models in production. The shift, analysts say, is reshaping how operators plan hardware and budgets.

A revised EU draft adds documentation and oversight duties for high-risk AI; firms in the bloc are weighing the timeline and compliance impact.

An updated draft framework introduces additional documentation and oversight duties for AI used in higher-risk settings. Companies operating in the bloc are weighing how the requirements would affect product timelines and compliance work.

The provider expanded the context window available to API customers, letting applications pass substantially more text in a single request. Developers see openings for document analysis and longer agent sessions, with cost and latency the open questions.

A bigger API context window lets apps send much more text per request — promising for document analysis and long agent runs, pending cost and latency.

The startup showed a neural processing unit built to run models at the edge under tight power budgets. It targets devices where sending data to the cloud is impractical, from sensors to industrial equipment.

A new low-power NPU runs models at the edge under tight power limits — aimed at sensors and industrial gear where cloud round-trips don't fit.

A consortium's new playbook walks enterprises through red-teaming AI before launch — prompt attacks, data leakage, and failure modes included.

An industry consortium published a practical guide for stress-testing AI systems before deployment, covering prompt attacks, data leakage, and failure modes. The playbook targets enterprises moving pilots toward production.