Coinbase fine-tuned open Qwen3.5-9B on internal fraud data, beating Opus 4.5 by up to 35 recall points

sebpaquet · x · 2026-10-13

Coinbase CEO Brian Armstrong shared that the company fine-tuned the open-weight Qwen3.5-9B on proprietary internal fraud data for a few weeks, and it now beats frontier models on fraud detection by up to 35 recall points, at far lower cost and latency.

Key takeaways:

Proprietary data is the decisive edge: fraud patterns are company-private and never touched the public internet, so no amount of pretraining compute helps — Opus 4.5 had seen zero real Coinbase fraud, while the 9B model trained on millions of internal labeled examples.
Speed and cost win too: the fine-tuned model runs on a single GPU and responds in 0.68s vs 1.5-3.3s for Opus, which matters since fraud checks run inline in the checkout flow.
Internalized workflow: Coinbase is making fine-tuning self-serve for employees, with an internal skill like /fine-tune-and-serve-llm to spin up and serve models.

Armstrong expects specialist models to increasingly beat frontier generalists on narrow tasks across companies.

Original post →

Coinbase fine-tuned open Qwen3.5-9B on internal fraud data, beating Opus 4.5 by up to 35 recall points

More from Companies & People