Atlassian says smaller moderation model cuts latency and increases throughput

Atlassian says it has deployed a fine-tuned 4B model ahead of its larger content moderation model, reducing end-to-end latency from 740ms to 350ms and increasing throughput by 42%. The system checks requests flowing through products including Rovo Chat and Rovo Search. Built with NVIDIA, the smaller model handles most requests while the larger model examines suspicious cases. Training used synthetic policy examples, adversarial variations and reasoning traces, allowing the smaller model to classify requests without generating reasoning at inference. Atlassian reports system precision rose from 94.0% to 98.1%, while recall rose from 50.9% to 56.6%. The smaller model's standalone recall of 75.5% is a separate evaluation result.