Why Our Own Models?

Why Our Own Models?

Most AI platform companies are resellers. They route your requests to someone else's models and add a management layer on top. ScaiLabs does that too — ScaiGrid connects you to any model from any provider — but we also train our own.

Why? Because there are tasks where a purpose-built model outperforms a general-purpose giant. A 13B model trained specifically for API interaction will handle a Swagger definition more reliably than a 70B generalist. A model optimized for speed handles routing decisions faster than one optimized for depth. And a model small enough for embedded systems opens deployment options that cloud-only models can't touch.

The Poolnoodle family is the result of our research in model training, fine-tuning, and efficiency. Each model is designed for a specific role in the ScaiLabs ecosystem, and they're trained to work together — handing off tasks to the right specialist when a different capability is needed.

Working Together

The Poolnoodle family isn't just a collection of independent models — they're designed as a collaborative system. FixerNoodle routes requests to the right specialist. BabyNoodle escalates to bigger models when it needs to. ToolNoodle gives the others hands to interact with external APIs. LongNoodle handles the documents that exceed everyone else's context window.

This collaborative architecture is a natural fit for ScaiCore, where Poolnoodle models can serve as the LLM backends for agents within a team — each Core using the model best suited to its role.

What's Next: EpicNoodle

We're actively researching the next evolution of the Poolnoodle architecture: a cognitive system that orchestrates multiple specialized models, adaptive memory, and tool libraries into a unified intelligence.

Rather than building ever-larger monolithic models, EpicNoodle takes the collaborative approach of the current family and formalizes it into an architecture where knowledge (retrieval and memory), capabilities (tools and logic), and reasoning (LLM attention) are explicitly separated and dynamically composed.

EpicNoodle is under active development as a ScaiCore-native architecture. Follow our research page for updates.

→ Learn about our research

---