Astra AI unveils Prism-XL: a 1.2T multimodal model optimized for on-device creative apps

AI · 5 min read

Astra AI today announced Prism-XL, a 1.2 trillion parameter multimodal model aimed squarely at creative and design applications. Unlike server-only giants, Prism-XL ships with a companion condensed runtime that offloads heavy tensor math to optional cloud accelerators while keeping critical layers local for low-latency interactions.

Early benchmarks shared by Astra show sub-50ms response times for standard image-editing prompts (e.g., style transfer and localized recoloring) on a modern ARM laptop when the runtime leverages quantized weights. The company highlighted progressive interaction patterns—edit, refine, undo—instead of single-shot generation, arguing this better matches designer workflows.

Prism-XL also introduces an API contract tailored for design tooling: deterministic undo tokens, delta-only payloads for collaborative edits, and a privacy-focused “ephemeral context” mode that discards user data on-device after inference. Astra positions the model as a plug-in for existing design suites rather than a walled garden product.