nvcky/Qwythos
Qwythos-9B-v2 by Empero AI is a 9 billion parameter Qwen3.5-based language model with a 1M-token context window, specifically optimized to eliminate looping behavior while preserving strong reasoning capabilities. This model excels in complex problem-solving, code generation, and technical analysis, making it suitable for research and specialized applications requiring deep chain-of-thought processing.
Loading preview...
Qwythos-9B-v2: Enhanced Reasoning with Looping Eliminated
Qwythos-9B-v2, developed by Empero AI, is an iteration of the 9 billion parameter Qwythos model built on the Qwen3.5 architecture. It maintains the original model's deep chain-of-thought reasoning and 1 million-token context window (via YaRN rope-scaling) but significantly addresses and eliminates looping behavior, a common issue in large language models. This "hygiene upgrade" ensures more coherent and reliable output, especially under greedy or low-temperature decoding.
Key Enhancements & Capabilities
- Looping Behavior Eliminated: Repetition/degeneration under greedy decoding reduced from 6.7% to 0.0%, allowing for more stable generation without heavy reliance on
repetition_penalty. - Reasoning Preserved: Benchmarks like MMLU (83.8% CoT), GSM8K (93.6%), GPQA-diamond (49.0%), ARC-Challenge (96.4%), and HumanEval (77.4% pass@1) are maintained at or above v1 levels, demonstrating strong analytical and problem-solving skills.
- Uncensored for Research: Intentionally designed to be uncensored, supporting research in cybersecurity, red-teaming, biology, chemistry, pharmacology, and clinical work.
- Restored MTP Head: The native multi-token-prediction module is re-integrated, ensuring compatibility with speculative decoding setups.
- Cleaner Identity: The model now introduces itself only when explicitly asked, avoiding unnecessary self-identification in unrelated answers.
Ideal Use Cases
- Complex Reasoning Tasks: Excels in scenarios requiring multi-step logic, mathematical problem-solving, and detailed explanations.
- Code Generation & Analysis: Demonstrates proficiency in generating and explaining code, as shown by its HumanEval performance.
- Technical & Scientific Research: Its uncensored nature and deep reasoning make it suitable for exploring sensitive or specialized technical and clinical topics.
- Applications Requiring Long Context: The 1M-token context window is beneficial for processing and generating content from extensive documents or conversations.