Post-Training a 35B Open Model for Code Search: 100x Cheaper Than Frontier Models

ypatil125 · x · 2026-09-05

Applied Compute and turbopuffer published a technical writeup on making small open-weight models competitive at large-scale code search: they built an RL environment to post-train Qwen3.6-35B-A3B to use search tools over an index of 9,000 real GitHub repos, with narrow (deep, ≤1,000-repo corpora) and open-ended (broad) task types. At 300 repos the approach is 3x faster than filesystem + grep and cuts marginal search cost by up to 100x vs frontier models, topping the needle-in-a-haystack task — showing that targeted RL post-training plus a dedicated index beats frontier models' ls/grep habits at scale.

Related event: RL-Trained 35B Open Model Cuts Code Search Costs 100x(2 posts)→

Original post →

More from coding & agent

coding & agent channel →