A new way of debugging open-weight models

On Friday, a group of more than 20 tech firms including IBM, NVIDIA and Microsoft published an open letter urging American lawmakers not to prematurely restrict so-called “open weight” AI models. Among the virtues of open-weight models, the authors write, is that they “give customers greater control” and allow organizations to “evaluate and adapt models to their own needs.” But what exactly what might that mean? How do open-weight models enable this extra degree of control?

A recent project from IBM helps illustrate the principle. vLLM-Hook is a plugin library that allows developers to program the inner workings of an AI model running on the vLLM inference engine. Its creators, of which GitHub lists a half dozen, say it lets users “inspect, analyze and steer the internal operations of large language models.”

Crucially, though, vLLM-Hook can only work on open-weight models like DeepSeek’s, Mistral’s or IBM’s. (While traditional software is either “open-source” or “closed-source,” AI models are not made of source code but rather huge arrays of numbers, which weigh the strength of connections inside the model.) Although proprietary “closed-weight” models like those from OpenAI and Anthropic get the lion’s share of attention, many open-weight models perform nearly as well: the independent model-testing firm Artificial Analysis reports that 84 of the highest-performing 150 LLMs are open-weight; one of these (a model from Chinese firm Z.ai) cracks the top ten.

It’s the openness that allows the control and customization of vLLM-Hook. Irene Ko, an IBM Research Scientist and one of vLLM-Hook’s developers, described a use case in an interview with IBM Think. Let’s say you’re the CTO of BankingCo, and you’ve customized a version of open-weight IBM Granite 4.1 to help your employees recommend BankingCo products to clients. And let’s say that 99% of the time, it works beautifully, but 1% of the time it erroneously identifies a rival bank’s similarly named product instead.

In practice, you’d upload instances of the erroneous output into vLLM-Hook and, thanks to the transparency afforded by open-weight models, the tool would then run its analysis, diagnosing exactly which connections in the model seem to be causing the error. Further, it can apply a fix moving forward, ideally zapping that last erroneous 1%.

“It’s a fancy way to debug an LLM, in such a way that there’s almost no additional cost,” Ko said. So far, vLLM-Hook’s makers count several businesses among its users, including AI startup KRNEL. The vLLM-Hook repo is currently available on GitHub.

Want to learn more about the inner composition of AI models (and the mysteries of their “weights”)? See our jargon-free explainer, “What does AI look like?”