In the news
Post-training open-weight models for large-scale code search
turbopuffer · Published · 3 min read
In 30 seconds
- What happened
- Applied Compute and turbopuffer trained a specialized 35B model for code search using reinforcement learning on two task types across 9,000 repositories.
- Why it matters
- Engineers building large-scale code search systems where grep and filesystem tools become expensive and slow on massive codebases.
- Watch out
- The fine-tuned model excels at narrow searches but does not yet match frontier models on open-ended pattern discovery tasks.
- post-train
- latency
- open-weight
The patterns behind this
- Structure-Aware Codebase Retrieval (Repo Map)
- Filesystem as Context (Context Offloading)
- Reinforcement Learning from Human Feedback
Each one covers how the technique works, when it earns its cost, and where it breaks.
The Agent Architect
One pattern, one tradeoff, one production failure story. A short weekly briefing for people building agentic systems.
Weekly email, one-click unsubscribe. We only use your address to send the briefing.