clglavan/magos-k8s-0.6b

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:0.8BQuant:BF16Ctx Length:32kPublished:May 24, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

The clglavan/magos-k8s-0.6b is a 0.6 billion parameter Kubernetes debugging assistant model, fine-tuned from Qwen3-0.6B. It specializes in generating concrete, executable commands and YAML manifests for Kubernetes, making it suitable for autonomous devops agents. The model is trained on Kubernetes documentation, API references, kubectl commands, and Prometheus alert runbooks, with a focus on reducing hallucinations and improving accuracy for agent-usable outputs.

Loading preview...

Model Overview

clglavan/magos-k8s-0.6b is a specialized 0.6 billion parameter Kubernetes debugging assistant, fine-tuned from the Qwen3-0.6B base model. Its primary goal is to provide concrete, executable commands and YAML manifests, making it ideal for integration into autonomous devops agents. The model has undergone a two-stage training process, including continued pre-training and supervised fine-tuning on a comprehensive corpus of Kubernetes documentation, API references, kubectl commands, and Prometheus alert runbooks.

Key Enhancements in v8

Version 8 introduces significant improvements over v7, including:

  • Increased Training Examples: ~6,740 Stage 2 training examples (+10%).
  • Schema-Filtered YAML: 521 YAML examples validated against K8s v1.34 OpenAPI spec, reducing invented fields.
  • Anti-Hallucination Contrast Bucket: ~317 new examples specifically targeting common kubectl flag and YAML field hallucinations.
  • General-Instruct Mix: ~600 Alpaca examples blended in to preserve base reasoning capabilities.

Performance Benchmarks

Evaluated by three independent judges across 25 prompts, v8 shows notable gains:

  • kubectl/CLI accuracy: Increased from 10 (v7) to 14 (+4 points).
  • YAML manifest validity: Improved from 11 (v7) to 12 (+1 point).
  • General reasoning: Recovered to 15 (+3 points from v7's 12).

What it's good at

  • kubectl command construction: Generates accurate commands with correct flags.
  • YAML manifest generation: Produces valid Pod, Deployment, Service, and other K8s resource manifests.
  • Diagnosing pasted errors: Analyzes kubectl describe output, logs, and alerts to suggest root causes and next steps.
  • Prometheus alert handling: Provides meaning and diagnostic steps for common Prometheus alerts.
  • Agent-style outputs: Delivers short, command-first responses suitable for automated execution.
  • Basic general reasoning: Retains foundational reasoning thanks to the Alpaca mix.

Limitations

  • Not designed for multi-step planning or complex tool chains.
  • May still hallucinate subtle or rare kubectl flags.
  • Accuracy decreases with multi-flag combinations on the same command.
  • Knowledge is limited to features released before mid-2026 source docs.

Finetunes
1 models