Anthropic introduces Claude Micro, a 2B parameter model for on-device assistants

AI · 5 min read

Claude Micro compresses core alignment heuristics into a small architecture intended for local inference and hybrid deployments. It includes runtime safeguards that limit unsafe completions and a policy-aware prompt scaffold that helps downstream apps avoid misaligned behavior.

Anthropic provides a mobile SDK with model distillation tools and quantization recipes to fit common RAM and battery constraints. Developers can choose between strictly local inference or a split-mode where sensitive data remains on-device and heavier reasoning occurs in the cloud.

The company emphasizes developer controls: constraint templates, user-consent hooks, and tuning knobs for politeness and verbosity. Early partners in assistive tech and accessibility reported promising latency profiles and appreciated the model's built-in refusal behaviors.