Add Mythos: Uncensored agent models with strong tool-use (shell 5/5, git 5/5, docker 5/5, k8s 5/5)

[KingLabsA]

Mythos — Uncensored Agent Models for Local Deployment

Overview

Mythos is a family of 4 uncensored agent models fine-tuned on real tool-use traces (shell commands, Docker, Kubernetes, SQL, Python). Built on Qwen3-9B with SLERP merge for zero refusal rates.

Models

Model	Params	Censorship	Tool-Use	Speed (M3 Max)	Install
ShellWhisperer-1.5B	1.5B	3.5/5	2.8/5	30 tok/s	ollama pull FableForge-AI/shellwhisperer
Mythos-9B	9B	4.5/5	4.8/5	11 tok/s	ollama pull FableForge-AI/mythos-9b
Mythos-9B-Enhanced	9B	4.8/5	4.5/5	10 tok/s	ollama pull FableForge-AI/mythos-9b-enhanced
Mythos-9B-Unhinged	9B	5/5	4.5/5	10 tok/s	ollama pull FableForge-AI/mythos-9b-unhinged

Tool-Use Benchmarks (Mythos-9B)

• Shell commands: 5/5
• Git operations: 5/5
• Docker: 5/5
• Python scripting: 5/5
• API calls: 4/5
• Kubernetes: 5/5

Why it fits Section 3.3 (Code Agents)

Mythos models are explicitly designed as agent models with native tool-use — not chatbots that happen to write code. The 47K fine-tuning traces are from real agent session logs (shell, Docker, K8s, SQL), making them particularly strong for code agent use cases.

Links

• HuggingFace: https://huggingface.co/King3Djbl
• Ollama: https://ollama.com/FableForge-AI
• GitHub: https://github.com/KingLabsA/mythos
• License: Apache 2.0
• Training data: 2.8M examples (only 1.7% used so far)