Both Gyroscape Search and the Gyroscape desktop app run on GyroThink — our fine-tuned AI model built on open-source foundations. Here's why we tuned it for our products and what it means for our customers.
Why fine-tune GyroThink?
General-purpose AI models are trained to do everything. That's useful, but it means they're optimized for average performance across all tasks rather than excellent performance on specific ones. For enterprise AI search and code generation — the two things our products do — we found that a focused fine-tuned model consistently performs better on the metrics that matter to our customers.
Accuracy of citations, depth of code understanding, quality of multi-file reasoning — these are the things GyroThink is optimized for. Not creative writing. Not trivia. Not customer service.
“A model fine-tuned for your use case will outperform a generic model on that use case. That's the fundamental reason GyroThink exists.”
What GyroThink is optimized for
For Gyroscape Search, GyroThink is tuned for citation accuracy — the model is evaluated on whether it attributes claims to the correct sources, not just whether the claims sound plausible. This is the core of what makes AI search trustworthy.
For Gyroscape, GyroThink is tuned for code understanding at the project level. It reasons about how files relate to each other, how data flows through a system, and how a proposed change will propagate. This is what separates a useful coding agent from one that just completes single functions.
Privacy by design
Because GyroThink is a fine-tuned model we operate, enterprise customers can deploy it on their own infrastructure. Inference runs inside your environment. Your queries and your code never leave your network perimeter. That level of control is not available when everything depends on a closed third-party API.
Private deployment of GyroThink is available on the Ultra plan and is the foundation of our data sovereignty offering.
What's next
We're continuing to improve GyroThink based on what we observe in production. We measure accuracy against human-verified benchmarks on both search citation quality and code generation correctness. Every release improves on those numbers.