Coinbase fine-tunes open-weight Qwen model for Onramp fraud prevention
For the fraud head at a regional bank
Coinbase CEO Brian Armstrong said on Oct. 12 that the company fine-tuned the open-weight Qwen3.5-9B model on proprietary internal data for fraud prevention in Onramp, according to a CoinGabbar report on his X post.1
Onramp lets people buy crypto in partner apps, including guest checkout without an account, where users have little history, so signals must come from recent transactions, CoinGabbar reported.1 An existing machine-learning model and rules engine screen payments first, then an AI agent reviews selected transactions with recent activity as context and returns a compact risk level; it can block a payment the first layer allowed but cannot override an earlier block.1
In an internal benchmark, the tuned Qwen system beat Opus 4.5 by 7.5 points in precision, 12.0 points in recall and 35.4 points in dollar-weighted recall, according to the report; the gains are percentage-point differences.1
Median latency in live traffic fell 55% compared with Opus 4.5; the report lists P50 latency at 0.683 seconds for Qwen3.5-9B and 1.515 seconds for Opus 4.5.1
Training ran on one NVIDIA RTX PRO 6000 96GB GPU for under $100, using reinforcement learning with verifiable rewards and LoRA, and serving used L40S GPUs scaling between 2 and 8 replicas, according to the report.1
The report cites an earlier online test using Claude Opus that showed 30% fewer fraudulent transactions and 22% less fraud value; the benchmark figures come from Coinbase's internal testing and the report does not say they were independently verified.1
Coinbase reports that a small model trained in-house on its own data beat a frontier model on a narrow fraud task at lower latency and training cost. Banks and fintechs weighing API-based frontier models for payment risk review may want to test tuned open-weight models against them on their own data, keeping in mind these results are internal and unverified.
CoinGabbar listed as developments to track whether the approach reaches other payment risk reviews and backend services, and further detail on how widely it will be deployed.1
Our take. A smaller model trained on your own fraud data may beat an expensive general model, but Coinbase’s internal benchmark is not proof for your payment flows. Require independent tests of recall, dollar-weighted recall, drift and block decisions before changing screening decisions.
Action. Model risk: Set independent tests for fraud-model recall, dollar-weighted recall, drift thresholds and override rates before approving a fine-tuned model.
What our records say
Claude's posture, per the vendor profile, is that Anthropic sells banks Claude seats and API access directly and through cloud marketplaces, and in 2026 added embedded engineering with Goldman Sachs and FIS, restricted cyber-defence access to its Mythos model for a few large banks, and a services firm co-founded with Goldman Sachs.1
From our records
- Claude (Anthropic)vendor · banks using it (8) Goldman, JPMorgan, BankUnited · alternatives Microsoft Copilot, ChatGPT Enterprise, Gemini Enterprise
Sources
- ReportCoinbase Fraud Prevention: Can a 9B Qwen Model Beat Opus? Coingabbar