Brian Armstrong says Coinbase fine tuned Qwen3.5 9B on internal data and it beat frontier models at fraud prevention
We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case). I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves. We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them: "/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started" I imagine we'll have some backend services that could be replaced entirely by this new architecture.
Context
The CEO of a major public company is making the case that small open weight models trained on proprietary data can beat expensive frontier models on narrow tasks, a pattern that could reshape how companies buy AI.
Coinbase published a series of technical posts from October 6 to 8 about using AI to screen fraud on Onramp, its service for buying crypto through partner apps. One post found that newer versions of several frontier models caught less fraud than their predecessors in its benchmark. Qwen3.5 9B is an open weight model from Alibaba's Qwen team.
Deep dive
On October 11, Coinbase CEO Brian Armstrong posted that the company fine tuned an open weight model, Qwen3.5 9B, on proprietary internal data and that it outperformed frontier models at much lower cost and higher speed on a specialized task, fraud prevention.
Armstrong said he expects more of this pattern inside companies: frontier models are amazing generalists, but specialist models will increasingly outperform on narrower tasks. He said Coinbase is making it easy for employees to fine tune models, including a skill in its internal harness, and that some backend services could be replaced entirely by this approach.
The charts he shared come from Coinbase's own October 8 technical post. According to them, the post trained model beat an Opus 4.5 decision agent by 7.5 percentage points in precision, 12.0 in recall, 9.6 in F1 and 35.4 in dollar weighted recall. Median latency was 0.683 seconds versus 1.515 for Opus 4.5. Coverage notes that detection and latency were measured in separate evaluations, and that the benchmark data is private, so outside researchers cannot replicate it.
Coinbase's earlier October 7 benchmark, according to CryptoSlate, found that newer Opus, Sonnet and GPT versions caught less fraud than the versions they replaced.
What to expect next
Watch whether Coinbase moves the fine tuned model into broader production and replaces backend services with specialist models as Armstrong suggested.
Also watch whether other large companies publish similar results, which would support the argument that open weight models plus proprietary data can beat frontier models on narrow tasks, and how frontier labs respond with their own fine tuning offers.