--- title: README emoji: 🍪 colorFrom: yellow colorTo: gray sdk: static pinned: false ---
Kukie logo # Kukie AI **Small open models for Kubernetes diagnosis** Kukie builds models that investigate a Kubernetes cluster with read-only tools and point to the root cause of an incident, so diagnosis can run on your own GPU without sending cluster data to an external model provider. [Website](https://durokukie.com/) · [Models](https://huggingface.co/Kukie-AI/models) · [Datasets](https://huggingface.co/Kukie-AI/datasets)
## Releases **[Kause-K8s-27B-FP8](https://huggingface.co/Kukie-AI/Kause-K8s-27B-FP8)** · model · v0.2
Finds the root cause of a broken Kubernetes workload with read-only tools. Qwen3.8-27B fine-tune, FP8, one 48 GB GPU with vLLM.
Unseen public problems: **52.3% → 58.3%** root-cause accuracy. **[Kause-K8s-train-v0.2](https://huggingface.co/datasets/Kukie-AI/Kause-K8s-train-v0.2)** · dataset
The 551 tool-calling conversations used to train v0.2. 47 fault types, Apache 2.0. ## What we're building - **Kubernetes diagnosis model** — a small open model that takes over read-only diagnosis for the Kukie agent, calling cluster tools and returning the cause as structured output. - **Evaluation harness** — disposable kind clusters with injected faults, used to score diagnosis on public and held-out scenario sets. - **Next:** v0.3 (training round 3) adds the fault types v0.2 saw least, such as admission webhooks, quotas, and DNS. --- This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP), funded by the Ministry of Science and ICT (MSIT), Korea, in 2026 (IITP-2026-AI·SW Maestro).