Inside Sky XT7: how we trained a 27B-parameter LLM on EU infrastructure

We get asked the same question on every enterprise call: "is Sky XT7 actually your model, or is it a wrapper?" Fair question — the AI market is drowning in vendors who repaint OpenAI APIs and call them "proprietary". This post walks through the training pipeline, the data sources, the safety stack and the inference setup that backs every product we ship.

Sky XT7 is a 27B-parameter dense decoder, trained from scratch on a curated corpus that includes high-quality German technical and legal text, Turkish business documents, and our own internal codebases. The base run used 384 H100s in Falkenstein over 11 days. We did not start from any third-party checkpoint.

The safety layer combines RLHF on operator-graded conversations, an in-house red-team agent that runs 24/7 against new checkpoints, and a content-policy classifier we shipped separately as part of Vault OTP. Output evaluation goes through three judges — one of which is allowed to be a frontier model — with majority vote required to promote.

• 

• 

Why we built our own model instead of fine-tuning someone else's

Two reasons. First, control: enterprise customers in DACH need to see the training pipeline, the data lineage, and the kill-switches. Wrapping a third-party API gives us none of that. Second, economics: at our scale of inference, owning the model and the silicon is cheaper than per-token billing within six months. The capex was justified.