Exchange Ranking
Top Exchanges
24h Volume Ranking
Popularity Ranking
Exchange BTC Balance
Proof of Reserves
Decentralized Exchanges
Funding Rate
Funding Heatmap
Liquidation Data
Max Pain
Long/Short Ratio
Whale L/S Ratio
Binance/Okex/Huobi L/S
Bitfinex Margin L/S
ETF Tracker
Solana ETF
XRP ETF
Hong Kong ETF
Bitcoin Treasuries
Crypto Reversal
Ethereum Reserves
HyperLiquid Wallet Analysis
Hyperliquid Whale Watch
Large Transactions
On-chain Movement
Bitcoin ROI
Stablecoin Market Cap
Options Analysis
News
Articles
Economic Calendar
Features
Wallet
Contract Calculator
Security
Collections
Watchlist
FollowingBrian Armstrong shares fine-tuned Qwen results
On October 12, 2026, CEO Brian Armstrong said that the company used proprietary internal data to fine-tune the open source weighting AI system Qwen3.5- 9B. According to its release, the model surpasses cutting-edge systems at a lower cost and at a higher speed on specific tasks. This task is Coinbase's fraud prevention.
This announcement adds new content to today's exchange's continued encrypted news stream using AI. Armstrong expects more companies to build narrow professional systems rather than relying solely on general-purpose systems.
The inner workings of Coinbase fraud prevention
The system is located within Onramp, which allows users to purchase cryptocurrencies in collaborative apps, including guest checkouts without an account. Because of the small visitor history, the signal must come from recent transactions.
Existing machine learning and rules engines first filter payment requests. The AI agent then reviews the selected transactions in conjunction with recent activities and returns a tight risk level. It can block transactions for which payments are allowed at the first level, but cannot overturn early interception decisions.
Previous online tests using Claude Opus showed a 30% reduction in fraudulent transactions and a 22% reduction in fraudulent value. Expected savings are approximately three times the cost of reasoning
Source: Official X Post
Qwen3.5- 9B vs Opus 4.5: Overview of benchmark results
在内部基准测试中，微调后的 Qwen 系统在四项检测指标上均优于 Opus 4.5。
指标 | 相比 Opus 4.5 的提升
精确度 (Precision) | +7.5 点
召回率 (Recall) | +12.0 点
F1 分数 | +9.6 点
美元加权召回率 (Dollar-weighted recall) | +35.4 点
在此场景下，美元加权召回率最为重要，因为它显示了捕获的欺诈价值量，而不仅仅是案件数量。数值为百分点差异，而非相对增益。
微调系统的延迟与训练成本
测量值 | Qwen3.5-9B | Opus 4.5
P50 延迟 | 0.683 秒 | 1.515 秒
P99 延迟 | 1.036 秒 | 3.312 秒
与实际流量相比，Opus 4.5 的中位延迟下降了 55%。训练在一块 NVIDIA RTXPRO 6000 96GB GPU 上进行，耗时不到 100 美元，使用了可验证奖励的强化学习和 LoRA（一种训练少量附加参数的方法）。服务部署使用了 L40S GPU，副本数在 2 到 8 之间扩展。
拥有开源权重模型如何改变经济结构
Armstrong believes that cutting-edge systems are generalists, while experts can win on narrow tasks. He also said that certain back-end services could be completely replaced by this architecture. Having weight brings several practical advantages to Coinbase's fraud prevention:
- Control: Select training data, checkpoints, and upgrade times
- Privacy: Distance transaction context from external APIs
- Reliability: Versioned weights, prompts and decision strategies
- Costs: Small trimming systems can replace more expensive general-purpose systems
Results 'implications for cryptocurrency exchange AI
Results suggest that Coinbase's fraud prevention may favor using small trimming systems rather than cutting-edge systems that rely solely on APIs. Attack patterns are constantly changing, so quality still depends on continuous evaluation and retraining, rather than a single good checkpoint.
Key developments worthy of attention include:
- Does this approach extend to other payment risk reviews and backend services
- New benchmark results because earlier internal testing showed that Updated cutting-edge version scores lower in recall and F1 scores
- More details on the scope of deployment of the method
For readers who follow cryptocurrency exchange news and cryptocurrency AI, this is a concrete example of how specialization trumps scale on a single narrow task. Coinbase's fraud prevention is a case worthy of attention.
