ResolveAI has released SRE-1, a 27B parameter model that beats frontier models in production debugging on a cost-accuracy basis. The model was built to dominate the Pareto frontier of accuracy and cost within this coding-adjacent domain.
A team of 11 people developed SRE-1 using open weights models and domain expertise, with a final training run of 7.4 days on 512 GPUs. The project was a collaboration with the Cortex training team at Snowflake.
