Both Gyroscape Search and GyroCode run on Gyrothink — the AI model we built for our own products. Here's why we built it and what it means for our customers.
Why build your own model?
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 model purpose-built for those tasks consistently outperforms general models 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 purpose-built for your use case will outperform a general model on that use case. That's the fundamental reason we built Gyrothink.”
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 GyroCode, 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 our own model, enterprise customers can deploy it on their own infrastructure. Inference runs inside your environment. Your queries and your code never leave your network perimeter. This is not possible with models you license from a third-party API — there, the inference always happens on someone else's server.
Private deployment of Gyrothink is available on the Enterprise 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.