The field teams of digital-payments-solution platform Innoviti follow a set procedure. 

When they visit large format retail stores to service payment terminals, they “take photos of themselves with the company logo and terminals, and even capture handwritten job sheets stamped and signed by the store manager”, explained Girish Varadarajan, chief data and AI officer at Innoviti. All as proof of work. But the whole process is cumbersome, involving massive amounts of visual data.

In early September, Innoviti upgraded to Qwen’s latest 3.8 open-weight model. This version goes beyond multimodal image processing to carry out core software engineering tasks like code generation and testing. The decision came after just three days of accuracy tests, compared with legacy players, who often mull over such switches for more than a year. 

The test itself was simple: could the model detect if a photo was actually taken inside the said retail store, benchmarked against a set of internally labelled “right” and “wrong” identifiers. After three days of testing, the model demonstrated near-zero false positives and less than 5% false negatives. It was good to go. 

In August, Qwen actually overtookBloombergAlibaba AI models hit 3 billion downloads, passing Meta, Google both Meta and Google’s AI models in download volumes. In fact, it’s fast becoming the default choice among BFSI players beyond Innoviti. The stock brokerage Zerodha uses Qwen too, with its engineering teams self-hosting the model directly on their personal laptops for programming work. The company said no customer or financial data resides on these employee systems.

Open-weight models like these have their advantages. “Weights”, in this case, refer to the numerical values representing everything the neural network learned during training. Open-weight models make these values publicly available—without sharing either the source code or the training dataset. 

This kills two birds with one stone. Those making such models get to shield proprietary training information from their competitors, unlike open-source models that disclose everything. Meanwhile, those using them get more control over their data as they can run them on their own private servers. Such models are also cheaper than closed-source options, more customisable, and afford greater visibility and auditability into how the model works.