Own your intelligence.
Sayqal is our model line. We take open models and tune each one for one job in Ouroboros OS: cited answers and tool calls, injection screening, knowledge-graph links and search, in Arabic alongside English. Our aim is that customers own models tuned on their own documents and in their own language, running on hardware they own.
Six models, one job each.
Three are live in the Ouroboros OS MVP. Three more are trained: one passed its blind test and ships next, one is held back after its blind test, and one is waiting for its blind test. Every figure below is from our own internal tests.
The numbers, drawn to scale.
Sayqal Chat is the model employees talk to. Sayqal Guard screens for prompt injection in English and Arabic. Sayqal Graph decides which links go into the knowledge graph.
Read these honestly. The chat evaluation is English and in-distribution, and there is no Arabic chat score yet. A test on real documents is next.
Prompt-injection screening, English and Arabic.
In-world holdout
Out-of-world holdout
In Arabic
Tuned for this exact job.
One job, done precisely
Sayqal Chat is trained to cite every claim, to make tool calls correctly, and to say "Insufficient authorized context." when the documents don't hold the answer.
Open weights underneath
Sayqal Chat is built on Qwen3.6-35B-A3B (Apache-2.0), a mixture-of-experts model with 3B active parameters, served at 4 bits so it fits on one box.
A model-agnostic pipeline
Our tuning recipe doesn't depend on one base model: it has produced six models on six different open bases. That matters to customers who need a different base model for policy or procurement reasons.
A company that tunes a model on its own documents should own that model, and run it on its own hardware.
A model that knows your company.
Talk to us about tuning Sayqal on your own documents, inside your own building.