Coinbase said that the internally fine-tuned Qwen3.5-9B performed better on fraud prevention tasks, at lower cost and higher speed.
Article
Coinbase Fine-tuned with internal data Qwen3.5-9B, outperforming frontier models in fraud prevention
BlockBeats news,October 12，Coinbase CEO Brian Armstrong posted on social media that the company fine-tuned the open-source model Qwen3.5-9B using proprietary internal data. On the specialized task of fraud prevention, it outperformed frontier models while costing less and running faster. The relevant charts show that the post-trained model significantly led on metrics including precision, recall,F1 and amount-weighted recall; it also improved over both the unadapted version and the Opus 4.5 decision agent; its median online inference latency was approximately 4.5 lower than that of Opus 55%.
Armstrong believes that although frontier models are excellent general-purpose players, specialized fine-tuned models will increasingly prevail on narrower tasks, whether operating independently or collaboratively.Coinbase is simplifying the fine-tuning process for internal employees and has already provided one-click launch commands in internal tools; some backend services may eventually be fully replaced by this type of architecture.
Original link https://m.theblockbeats.info/flash/371257
Key points
01
BlockBeats said that October 12，Coinbase CEO Brian Armstrong published related content on social media.
02
According to Armstrong,Coinbase used proprietary internal data to fine-tune the open-source model Qwen3.5-9B。
03
Armstrong said that the fine-tuned model outperformed frontier models on the specialized fraud prevention task, while costing less and running faster.
04
The charts cited in the report show that the model outperformed the unadapted version in precision, recall,F1 and amount-weighted recall, as well as the Opus 4.5 decision agent.
05
The report said that the median online inference latency was approximately 4.5 lower than that of Opus 55%。
06
Coinbase is simplifying the fine-tuning process for employees and has already provided one-click launch commands in internal tools.
07
Armstrong believes that specialized fine-tuned models will increasingly prevail on narrower tasks, and said that some backend services may eventually be fully replaced by this type of architecture.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Fine-tuning can be understood as having a general-purpose model continue learning from a company's own data so that it can adapt to a specific task. Here, the specific task is fraud prevention. The reported results mean that Coinbase said that a smaller model adapted using internal data outperformed the comparison models on multiple evaluation metrics for this single task, while also having lower response latency. This is a comparison for a specific task and does not mean that the model is stronger in all capabilities.
Why it matters to readers
If these results continue to hold in real-world systems,Coinbase could run some fraud prevention workloads at lower cost and with shorter latency, and could also expand the use of internally fine-tuned models. However, the original text provides no verifiable cost data or actual business impact.
Beginners can focus on one core concept: the strongest general-purpose model does not necessarily perform best on every specialized task. Fine-tuning a smaller model with a company's own data may yield results and speed better suited to its business, but evaluation results still need to be assessed alongside the testing methods and actual deployment performance.
Risks and unknowns
The evidence does not provide Brian Armstrong's original social media post; the relevant claims are mainly based on secondary reporting.
The evidence does not disclose the testing method, sample size, data scope, or statistical significance, so the fairness of the comparison cannot be independently confirmed.
Related Developments
Loading event timeline…
“Lower cost” is not accompanied by specific figures or a calculation basis.
Apart from the approximately 55%median latency reduction, no specific figures are provided for the other metrics.
It remains unclear whether the model has been used in production and what its actual false-positive, false-negative, and business impact are.
Whether some backend services will be replaced by this type of architecture remains an uncertain future possibility.
Related concepts
Fine-tuning
This term is not in the glossary yet. Browse related concepts in the glossary.
