Pq234/Qwythos-9B-Claude-Mythos-5-1M
Qwythos-9B-Claude-Mythos-5-1M by Empero is a 9 billion parameter reasoning model built on a Qwen3.5-9B base, post-trained on over 500 million tokens of Claude Mythos and Fable traces. It features a 1,048,576-token context window via YaRN scaling, native function calling, and self-correction with tools. This model excels in technically demanding domains like cybersecurity, biomedical, and quantitative reasoning, offering significantly improved MMLU and GSM8K performance over its base.
Loading preview...
Overview
Empero's Qwythos-9B-Claude-Mythos-5-1M is a 9 billion parameter reasoning model, fine-tuned from a Qwen3.5-9B base. It was post-trained on over 500 million tokens of high-quality Claude Mythos and Fable traces, incorporating chain-of-thought generated by Empero AI's rethink tool. This model is designed for advanced reasoning tasks and is intentionally uncensored to engage with complex technical questions without refusal.
Key Capabilities
- 1,048,576-token context window: Achieved through YaRN rope-scaling, enabling whole-codebase reasoning, long agentic trajectories, and multi-document research.
- Enhanced Reasoning Performance: Demonstrates significant improvements over the base Qwen3.5-9B, with +34 pts on MMLU and +30 pts on gsm8k-strict.
- Native Function Calling: Supports OpenAI/Qwen3.5-style function calling out-of-the-box, allowing for tool use without additional fine-tuning.
- Self-Correction with Tools: Proven to self-correct and provide source-cited, factually correct answers when integrated with tools like Python executors and web search.
- Domain Expertise: Strong performance in cybersecurity, red-teaming methodology, biology, pharmacology, clinical medicine, and quantitative reasoning.
When to Use This Model
Qwythos-9B is ideal for applications requiring deep reasoning, long-context understanding, and reliable tool integration. Its uncensored nature makes it suitable for technically demanding questions in specialized fields where other models might refuse or provide boilerplate responses. It is particularly well-suited for agentic settings where factual verification and source citation are critical.