Coinbase fine-tunes Qwen3.5-9B for under $100, cuts Onramp fraud by 30%
- —Fraudulent Onramp transactions fell 30%, with dollar value of fraud down 22%
- —The 9B-parameter model beat Opus 4.5 on precision, recall, F1 and dollar-weighted recall
- —Median latency was 0.683 seconds for Qwen versus 1.515 seconds for Opus 4.5
- —Training cost under $100 on a single NVIDIA RTX PRO 6000 96GB GPU
Why it matters: The case shows narrow, cheap post-trained models can outperform expensive frontier systems in fraud detection, cutting costs for crypto services.
Source: Crypto Briefing