Coinbase fine-tuned Qwen3.5-9B for fraud prevention — beat frontier models, cheaper
lmoroney · x · 2026-10-14
Coinbase shared a useful data point for open-weight builders: the team fine-tuned Qwen3.5-9B on proprietary internal data for fraud prevention, and it outperformed frontier models on that narrow task while running faster and costing much less. CEO Brian Armstrong expects more companies to deploy specialist models for narrow jobs. Coinbase even built an internal skill so employees can kick off fine-tuning from a plain-language request like "I want to fine-tune Qwen for payments risk review." Tip: collect 200 real labeled examples this week and score your current model to see if a fine-tuned specialist is worth it.
More from coding & agent
- Claude builds its own test tool, fixes DreamPark MR sky-flicker bug in minutes — nptacek · 2026-10-14
- Argon arrives in Google Antigravity but users say it's useless even on the 5x plan — waqasalibrand · 2026-10-14
- CMU paper: a 4B model rewrites an agent's harness and beats the 35B model it learned from — irinarish · 2026-10-14
- Harness ≠ infrastructure: sandboxing and persistence belong to a different layer — blaizedsouza · 2026-10-14
- Anthropic publishes security guide for locking down Claude Managed Agents sandboxes — brada · 2026-10-14
- Dev used Gemini as software architect to build an Apple Watch caffeine tracker — pythononrailz · 2026-10-14