Coinbase fine-tuned Qwen3.5-9B on fraud data, beating Opus 4.5 at half the latency
J0se · x · 2026-10-12
Coinbase CEO Brian Armstrong says fine-tuning open-weight Qwen3.5-9B on proprietary fraud-prevention data outperformed frontier models at much lower cost: +35.4pp dollar-weighted recall vs an Opus 4.5 agent, with P50 LLM latency of 0.683s vs 1.515s (55% lower). Coinbase is making fine-tuning a one-command internal skill, expecting specialist models to replace some backend services.
More from Companies & People
- Security researcher reviews Meta Muse privacy design: 'even more encrypted' isn't a security standard — niloofar_mire · 2026-10-12
- Cambridge CRUK recruiting clinical PhD on spatial transcriptomics + AI for pancreatic cancer — mo_lotfollahi · 2026-10-12
- Sam Altman Gets the $1.4 Trillion Question: $13B Revenue vs $1.4T in Commitments — AlexTensor · 2026-10-12
- Engineer joins NVIDIA's Groq LPU compilers team, working on multi-chip partitioning — blelbach · 2026-10-12
- Toyota's Kaizen rule: never lay off workers saved by AI efficiency gains — dotey · 2026-10-12
- Ex-Apple engineer: two years in QA taught me to never treat quality as second-class — jdluk87 · 2026-10-12