Publishing the Orkhon Model Family
The June Orkhon updates published the project as a model family: a from-scratch PyTorch LLM stack with CI, demo, prompt-routing, and public-site wiring.
Umut Korkmaz
Published: 4 min read
On this page3
On June 16 and June 20 the Orkhon work moved from local experiment toward a public model-family story. The commits tell the shape of it: publish Orkhon, stabilize CI, publish the unified Orkhon model family, route common assistant prompts, and pin the demo Space dependency.
What Orkhon is
Orkhon is my from-scratch LLM stack in PyTorch. It is deliberately auditable: tokenizer, decoder-only Transformer, training loops, alignment experiments, evaluation, serving, and export paths live in the same family of work. The goal is not to pretend it is a frontier model. The goal is to make a small, inspectable language lab that I can run, explain, and extend myself.
Why the family framing matters
Calling it a family keeps the project honest. There can be tiny smoke-test models, demo models, data tools, evaluation artifacts, and website pieces without forcing all of them into one repo narrative. That is more useful than presenting one checkpoint as the whole product.
What shipped
The live site now points to the Orkhon page, the repo has the public publish path, CI was stabilized around the required dependencies, and the demo path was pinned so visitors do not hit a moving dependency target. It is the kind of infrastructure work that makes an open-source AI project feel real: not just a model file, but the system around it.