Tivaphraen/Geryon-9B-v1
Tivaphraen/Geryon-9B-v1 is an experimental 9 billion parameter DARE-TIES merge model built on the Qwen 3.5 architecture. It integrates specialized fine-tunes for general reasoning, agentic coding assistance, and long-form text generation, aiming to combine diverse capabilities without catastrophic forgetting. This model is optimized for complex problem-solving and structured output, making it suitable for advanced AI workflows.
Loading preview...
Geryon-9B-v1: An Experimental Merge for Advanced Reasoning and Coding
Geryon-9B-v1 is an experimental 9 billion parameter language model developed by Tivaphraen, leveraging the Qwen/Qwen3.5-9B dense architecture. Its core innovation lies in using the DARE-TIES merge method to combine multiple highly specialized fine-tunes, aiming to integrate diverse capabilities like reasoning, coding, and agentic workflows without experiencing catastrophic forgetting.
Key Capabilities
- Enhanced Reasoning: Incorporates a fine-tune trained on over 500M tokens of Claude Mythos/Fable traces, emphasizing Chain-of-Thought (CoT) for complex problem-solving.
- Agentic Coding: Integrates fine-tunes for agentic coding and tool use, mimicking Claude Fable 5 and GPT-5.5 terminal-agent styles, and includes a coding-agent model trained on 425K agentic coding trajectories.
- Long-Form Generation: Designed for robust long-form text generation across various domains.
- Multilingual Support: Inherits strong multilingual capabilities from its Qwen base, demonstrated through instruction following and generation in multiple languages.
Intended Use Cases
- Chat and Assistant Interactions: Ideal for general-purpose conversational AI.
- Coding Support: Excellent for generating and debugging code, with robust error handling and detailed comments.
- Reasoning-Heavy Tasks: Excels in logical problem-solving and structured output, such as JSON generation.
- Experimentation: A valuable model for researchers exploring merged specialist models and local inference workflows.
Performance Snapshot
While comprehensive benchmarking is ongoing, local testing on the Q8_0 GGUF quantization shows strong performance on GSM8K, with an exact match score of 0.8514 (flexible) and 0.8506 (strict). This indicates effective preservation of reasoning strength from its parent models, with added coding and agentic capabilities. The model demonstrates strong instruction following, robust Python coding, and strict JSON formatting capabilities, though some schema adherence may vary.