luxuansang/Qwythos-9B-v2
Qwythos-9B-v2 by Empero AI is a 9 billion parameter Qwen3.5-9B hybrid model with a 1 million token context window, specifically engineered to eliminate looping behavior while preserving strong reasoning capabilities. This model maintains its uncensored research posture and multimodal-capable architecture, making it suitable for complex analytical tasks, cybersecurity, and scientific research where robust, non-repetitive output is critical. It excels in deep chain-of-thought reasoning and technical problem-solving, offering a cleaner identity and restored multi-token-prediction head for improved speculative decoding.
Loading preview...
Qwythos-9B-v2: Enhanced Reasoning with Looping Eliminated
Empero AI's Qwythos-9B-v2 is an improved iteration of the 9 billion parameter Qwythos model, built on a Qwen3.5-9B hybrid architecture with a 1 million token context window (YaRN). The primary advancement in v2 is the elimination of looping behavior, reducing repetition/degeneration under greedy or low-temperature decoding from 6.7% to 0%. This allows for more reliable generation without heavy reliance on repetition_penalty.
Key Enhancements & Capabilities
- Looping Behavior Eliminated: Achieves 0.0% looping rate under greedy decoding and 0.7% at temperature 0.6, a significant improvement over its predecessor.
- Reasoning Preserved: Maintains or improves strong performance across key benchmarks including MMLU (83.8% CoT), ARC-Challenge (96.4%), GPQA-diamond (49.0%), GSM8K (93.6%), and HumanEval (77.4%).
- Restored MTP Head: The native multi-token-prediction module is re-integrated, ensuring configuration and weights align for speculative decoding setups.
- Cleaner Identity: The model now introduces itself only when explicitly asked, avoiding unnecessary self-identification in unrelated answers.
- Uncensored for Research: Intentionally designed to be uncensored for applications in research, cybersecurity, red-teaming, biology, chemistry, pharmacology, and clinical work.
- Long Context: Features a 1,048,576-token context window via YaRN rope-scaling.
- FTPO Training: Looping was fixed using Final-Token Preference Optimization (FTPO), a method that gently trains the model to prefer coherent alternatives at loop-start positions without impacting overall knowledge or reasoning.
Ideal Use Cases
- Complex Reasoning Tasks: Excels in scenarios requiring deep chain-of-thought, such as mathematical problems, logical puzzles, and code generation.
- Technical & Scientific Research: Its uncensored nature and robust reasoning make it suitable for sensitive technical, cybersecurity, and scientific inquiries.
- Applications Requiring Reliable Output: Ideal for systems where repetitive or degenerate output is unacceptable, such as automated content generation or interactive agents.
- Code Generation & Analysis: Demonstrates strong performance in coding tasks, as shown by its HumanEval score and ability to generate well-explained code solutions.