Tivaphraen/Geryon-9B-v1
Tivaphraen/Geryon-9B-v1 is an experimental 9 billion parameter DARE-TIES merge model built on the Qwen3.5-9B architecture, featuring a 32768-token context length. It integrates specialized fine-tunes for general reasoning, agentic coding, and long-form text generation, aiming to layer multiple capabilities without catastrophic forgetting. This model excels in complex problem-solving, code assistance, and structured output tasks, making it suitable for advanced agentic workflows and research into merged specialist models.
Loading preview...
Geryon-9B-v1: An Experimental DARE-TIES Merge Model
Geryon-9B-v1 is a 9 billion parameter experimental merge model, leveraging the Qwen/Qwen3.5-9B dense architecture with a 32768-token context length. Developed by Tivaphraen, its core objective is to explore the viability of combining multiple highly specialized fine-tunes into a single model while preserving their distinct capabilities and avoiding catastrophic forgetting. The merge integrates three key fine-tunes:
Key Capabilities
- Enhanced Reasoning: Incorporates
empero-ai/Qwythos-9B-Claude-Mythos-5-1M, fine-tuned on over 500M tokens of Claude Mythos traces, with a strong emphasis on Chain-of-Thought (CoT) reasoning. - Agentic Coding & Tool Use: Includes
empero-ai/Qwable-9B-Claude-Fable-5for agentic coding and reasoning, designed to imitate Claude Fable 5 and GPT-5.5 terminal-agent style tool use. - Robust Code Generation: Integrates
Tesslate/OmniCoder-9B, a coding-agent model fine-tuned via LoRA on 425K agentic coding trajectories. - Multilingual Support: Retains the Qwen base's inherent multilingual capabilities, demonstrated through strong instruction following in multiple languages.
- Structured Output: Shows proficiency in generating strict JSON and other structured formats, beneficial for agentic workflows.
Performance Highlights
Initial local testing on a Q8_0 GGUF quantization shows strong performance on GSM8K, with an exact match (strict) score of 0.8506, surpassing the base Qwen3.5-9B and even slightly exceeding the published strict score of its Qwythos parent. This suggests effective preservation of reasoning strength while adding coding and agentic features.
Good For
- General reasoning and complex problem-solving
- Coding assistance and robust Python script generation
- Agentic workflows requiring structured outputs (e.g., JSON)
- Long-form text generation
- Experimentation with merged specialist models and local inference research
Limitations
As an experimental merge, Geryon-9B-v1 may exhibit unexpected behaviors or inconsistent style transfer. It inherits an uncensored heritage from some parent models, meaning it may not reliably refuse sensitive or unsafe instructions. Users should implement appropriate safety filters and guardrails, especially given its experimental nature and the potential for hallucinations common in 9B-class models.