Whoops! Sorry about that.

TL;DR: Using SkyRL on Anyscale, we post-trained Qwen3.5-9B for Onramp fraud detection and served it with Ray Serve LLM and vLLM. On our benchmark, it exceeded Opus 4.5 across all four detection metrics, including 9.6% higher F1. In online inference, we reduced end to end latency by 55%.

TL;DR: Coinbase will power USDC as the default dollar stablecoin in the Samsung Wallet, alongside a licensed partner. USDC balances will be custodied by Coinbase Prime Vault

TL;DR: Newer LLMs are not necessarily better at domain-specific tasks. On our fraud benchmark, newer Opus, Sonnet, and GPT versions all had lower recall, F1, and dollar-weighted recall—even when precision improved. The lesson: choose models using high-quality, domain-specific benchmarks, with reproducible evaluation and ideally deterministic scoring against verified outcomes. These findings also motivated our next step: post-training open-weight models for greater control over quality, cost, and deployment.

การผสานรวม Deribit ของเราเสร็จสมบูรณ์แล้ว และเป็นครั้งแรกในประวัติศาสตร์ของตลาดที่ Coinbase Financial Markets จะเชื่อมต่ออนุพันธ์คริปโตของสหรัฐฯ และทั่วโลกเข้าด้วยกันในกลุ่มสภาพคล่องที่มีการกำกับดูแลเพียงแห่งเดียว

TL;DR: How Coinbase built, evaluated, and post-trained an Onramp fraud agent to achieve 9.6% higher performance (F1 score) and 55% lower end to end latency than frontier models.

TL;DR: We moved more than 300 Temporal namespaces, across more than 300 services, from a cluster we ran ourselves onto Temporal Cloud. You cannot finish the move by repointing services at Temporal Cloud while workflows are still running on the old one. Where teams could pause new work, we waited for in-flight runs to finish and then switched; for trading, payments, and long-running work, we ran the old cluster and Temporal Cloud, and moved new workflows over in steps, with rollback as a config change. Private connectivity, per-namespace service allowlists, and encrypted payloads were in place before the first production namespace moved.