ResolveAI Introduces SRE-1, a 27B Model That Matches Frontier Accuracy at Lower Cost for Production Incidents
·2026.11.20 08:38
Key point
The 27B model uses reinforcement learning to improve single-investigation pass rates from 6.9% to 38.8% on live production cases.
Details
- ResolveAI launched SRE-1, a 27B open-weights model trained specifically to identify root causes in production alerts, available exclusively through the company's platform.
- On 49 difficult live production cases, an ensemble of five SRE-1 investigations matches the strongest frontier model tested while costing less, with reinforcement learning boosting single-investigation pass rates from 6.9% to 38.8%.
- As a complete system, ResolveAI solved approximately twice as many incidents as general-purpose coding agents on 178 graded investigations, achieving root-cause scores 15 to 25% higher on the hardest third of alerts.
- Primary On-call, powered by SRE-1, is generally available and handles up to 80% of paged alerts without human escalation, with customers reporting a 10x improvement in mean time to resolution.
Model Development
- Built by 16 AI researchers over a 7.4-day training run using up to 512 GPUs.
- Training utilized reinforcement learning with tools pinned to the alert timestamp to prevent the model from accessing future data.
- Developed with support from Snowflake Cortex Training, with Vivek Raghunathan, SVP of Engineering at Snowflake, citing the team's modeling and RL talent.
Operational Impact
- Primary On-call investigates across telemetry, infrastructure, code, and operational knowledge to close alerts or escalate to engineers.
- The system connects related alerts to identify and fix underlying issues, preventing recurring incidents.
- Salesforce reported closing more than 70% of paging alerts without waking engineers during initial dark launch testing.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.