clglavan/magos-k8s-0.6b

TEXT GENERATIONConcurrency Cost:1Model Size:0.8BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:May 24, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

clglavan/magos-k8s-0.6b is a 0.6 billion parameter reasoning model derived from Qwen3-0.6B, specifically fine-tuned for Kubernetes diagnostics. It excels at mapping Kubernetes symptoms to corrective actions, generating structured reasoning traces, and providing precise kubectl/promtool commands or YAML patches. This model is optimized for use as an inner-loop reasoner in devops agents, focusing on event-grounded diagnostic reasoning and YAML manifest generation.

Loading preview...

Model Overview

magos-k8s-0.6b is a 0.6 billion parameter reasoning model, based on Qwen3-0.6B, specialized in Kubernetes diagnostics. It undergoes a two-stage full-weight training process: continued pre-training on Kubernetes documentation and API references, followed by supervised fine-tuning on event-to-YAML diagnostic pairs. The model generates structured <think> reasoning traces culminating in concise actions like kubectl commands, YAML patches, or root cause analyses.

Key Capabilities

  • Kubernetes Diagnostics: Correlates kubectl get events or kubectl describe output to responsible YAML fields and suggests fixes.
  • YAML Manifest Handling: Strong performance in generating and reviewing YAML manifests across various Kubernetes Kinds (Pod, Deployment, Service, etc.).
  • Prometheus Alert Processing: Provides meaning and diagnostic steps for prometheus-operator runbook alerts.
  • Structured Reasoning: Delivers short, templated reasoning traces that lead to concrete next actions, ideal for integration into devops agents.

What's New in v15

v15 represents a significant redesign, focusing on event-grounded diagnostic reasoning. It was trained on ~16.6k matched BROKEN/HEALTHY event→YAML diagnostic pairs, leading to improved performance in YAML handling and diagnosis. Benchmarks show v15 outperforming v8 overall, particularly in YAML (+9) and diagnosis (+3) buckets, while reducing hallucinations by 14% and increasing the agent-usable rate to 31%. A known regression is its performance on plain single kubectl commands, where it tends to over-diagnose.

Intended Use

This model is designed as a reasoning component for learning, prototyping, and local devops agents. It is recommended to run the model greedily with repetition_penalty = 1.0 for optimal structured output. Users should always verify generated commands or YAML against current Kubernetes documentation, as knowledge is frozen at the training snapshot (mid-2026).

Finetunes
1 models