A 27B decision model based on Qwen3.5 that produces typed decisions and probabilities from text, JSON, and images, with cached context support.
Morph describes Reflex as a feedback layer for continually learning agents. It labels conversation turns for signals such as user frustration, jailbreaks, looping, and policy violations, and says those outputs can feed evaluations, fine-tuning, and reinforcement-learning rewards. The page also describes training custom signals using labeled or synthetic data. Morph reports under 90 ms end-to-end latency and support for up to 64k context; these are claims on the product page, not independently verified measurements.
