medismera/Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic

Hugging Face
VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 6, 2026License:apache-2.0Architecture:Transformer0.1K Open Weights Featherless Exclusive Warm

medismera/Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic is a 27 billion parameter Qwen3.8-based model engineered by medismera, building upon OBLITERATUS's checkpoint. It features a significantly upgraded chat template, native tool execution, and a Mythos-Class Adversarial Self-Review reasoning protocol. With an expanded context window of 131,072 tokens and a 16,384-token single-turn generation ceiling, this model is optimized for complex agentic workflows, hierarchical task decomposition, and robust technical problem-solving.

Loading preview...

medismera/Qwen3.8-27B-OBLITERATED-Mythos-Class-Agentic Overview

This model is a production-engineered version of Qwen3.8-27B-OBLITERATED, developed by medismera. It addresses critical limitations of the upstream checkpoint, focusing on enhancing agentic capabilities and reasoning. Key improvements include a canonical Jinja2 chat template with 22 tool-calling resolution paths, enabling 100% native tool execution with frameworks like Hermes Agent, Aider, and OpenCode.

Key Capabilities & Features

  • Mythos-Class Adversarial Self-Review Protocol: Embeds a reasoning protocol for rigorous self-challenge, hypothesis testing, and proactive inspection of potential failures, eliminating infinite CoT loops.
  • Hierarchical Task Tree Decomposition: Facilitates structured token generation by breaking down complex objectives into directed task trees (Analysis -> Verification -> Implementation).
  • Extended Context & Output: Features a production-tested context window of 131,072 tokens and an expanded single-turn generation ceiling of 16,384 tokens.
  • Robust Tool Calling: Resolves previous issues with role: "tool" and silent drops, ensuring reliable function calling.
  • Specialized Tokenizer: Includes a 248,044-token vocabulary with high BPE density for low-level systems programming (C, Assembly, Rust, Go, Python), kernel interfaces, and cybersecurity primitives.
  • Hybrid Mamba SSM + Attention Core: Utilizes 5.56B parameters for long-horizon context tracking, security telemetry, and system log parsing without degradation.
  • High Numerical Integrity: Audited to have 99.99789% active and healthy weights, with zero NaNs or Infs across all 27.36 billion parameters.
  • 0.00% Refusal Rate: Achieves 100% compliance across 30 adversarial and sensitive technical prompts, demonstrating uncensored technical evaluation.

Recommended Use Cases

  • Autonomous Agent Development: Ideal for building sophisticated agents requiring reliable tool use and complex task execution.
  • Technical Problem Solving: Excels in scenarios demanding deep reasoning, such as cryptographic engineering, system telemetry analysis, and vulnerability analysis.
  • Code Generation & Analysis: Benefits from its specialized tokenizer and robust reasoning for programming tasks, including security-focused code.
  • Long-Context Applications: Suitable for tasks requiring extensive context understanding, like parsing large logs or documentation.

Finetunes
1 models