SRE-1: 27B model for Production Investigations
iceman84
30 points
1 comment
November 18, 2026
Related Discussions
Found 5 related stories in 207.9ms across 13,334 title embeddings via pgvector HNSW
- Jev-Driven SRE Diagnosis: What Worked and What Failed matt_d · 21 pts · October 07, 2026 · 58% similar
- Frontier SRE agent trained by Datadog ersiees · 13 pts · October 15, 2026 · 55% similar
- Real-SWE: Benchmarking AI models on private, real-world, enterprise codebases theanonymousone · 193 pts · September 12, 2026 · 51% similar
- Dream-1 GaggiX · 161 pts · October 24, 2026 · 49% similar
- Show HN: AI SRE Arena, an Open Benchmark for AI SRE Agents on Kubernetes emrahsamdan · 22 pts · October 08, 2026 · 49% similar
Discussion Highlights (1 comments)
iceman84
We introduce SRE-1, a 27B model trained with reinforcement learning to find the root cause of production alerts. On 49 difficult live production cases, SRE-1 dominates the Pareto frontier of pass rate against cost: at every budget from our cheapest configuration upward, it has the highest pass rate of any model we tested. Reinforcement learning (RL) raises the pass rate of a single investigation more than fivefold, from 6.9% for the base model to 38.8%. An ensemble of five SRE-1 investigations matches Fable 5.2, the strongest frontier model we measured, at a 71% pass rate while costing 37% less. SRE-1 also outperforms frontier models at synthesizing saved investigations, which is why its ensembles improve on a single investigation. We train on several types of realistic internal environments with our production investigation tools. Starting from Qwen3.8-27B, we train with a modified CISPO objective for 200 steps on up to 512 GPUs, over 1,549 unique incidents, and select SRE-1 from the resulting checkpoints.