Pharma companies should fine-tune open AI models on their own data, as Coinbase did for fraud detection, says Xaira’s Bo Wang

Pharma companies should fine-tune open AI models on their own data, as Coinbase did for fraud detection, says Xaira’s Bo Wang

Coinbase CEO Brian Armstrong wrote that the cryptocurrency exchange had fine-tuned the open model Qwen3.5-9B on internal data. It outperformed leading AI labs’ frontier models at fraud detection while running faster and costing less. An hour and a half later, on October 12, Bo Wang proposed the same approach for pharma companies.

Coinbase detects fraud in Onramp, a service for buying cryptocurrency within partner apps. Alongside its existing model and rules, the company introduced an AI agent built on Anthropic’s Claude Opus model to review transactions. In a live experiment, the group using the agent had 30% fewer fraudulent transactions than the control group.

On historical transaction data, the new Opus 5, Sonnet 5 and GPT-5.6 models detected a smaller proportion of fraud than their predecessors, with declines of 0,8–22,2 percentage points. With Opus 4.5, Coinbase said, its control “ended at the API boundary.” The API is the interface through which software accesses a model running on its developer’s servers. Coinbase could choose the prompts and model version, but could not adapt the weights, the numbers that encode what the model has learned, to Onramp’s fraud patterns.

Coinbase then turned to Qwen3.5-9B, an open model with nine billion weights. The weights are publicly available, allowing users to download the model, run it on their own infrastructure and fine-tune it on their own data. Coinbase used reinforcement learning: a formula compared the model’s transaction risk assessment with the known outcome and assigned a reward. Training ran on a single GPU and cost less than 100 dollars.

In Coinbase’s own evaluation, the fine-tuned model outperformed the Opus 4.5 agent. Precision, the proportion of flagged transactions that were actually fraudulent, was higher by 7,5 percentage points. Recall, the proportion of fraudulent transactions detected, was higher by 12,0 percentage points, and the proportion of the total fraudulent transaction value detected was higher by 35,4 percentage points. Coinbase says open weights let it decide what to train and when to update the model, while running it on its own infrastructure keeps transaction data out of external APIs. In return, the company must continually evaluate and retrain the model as fraud patterns change.

Wang, Xaira’s chief AI scientist and one of the creators of scGPT, a model for single-cell data, writes: “Cost and speed are the small prizes here, but sovereignty is the big one.” In his view, a pharma company that rents intelligence by paying to use a frontier lab’s model is strategically vulnerable: others determine access, pricing and development plans, while its most sensitive data crosses a boundary it does not control.

Every major pharma company, Wang continues, has an asset that AI labs cannot reproduce: decades of proprietary data from large-scale molecular screening, laboratory assays and failed trials. By fine-tuning open weights on these data, a company gains ownership of the weights, model deployment and intellectual property. He describes this as sovereignty at the company level. His prediction follows:

“The first pharma company to grasp this will shorten drug discovery timelines by years. The rest will end up renting intelligence about their own molecules.”

There is already a precedent in pharma: on September 14, Nature reported that five pharma companies had fine-tuned the open model OpenFold3, which predicts the three-dimensional structure of a protein together with a drug molecule, using 20 167 proprietary structures. According to a report from the Apheris platform, an evaluation on 1 056 structures from the same companies that had been excluded from training found high-quality protein–ligand interfaces in 52,1% of the new version’s predictions, compared with 35,6% for the original OpenFold3.