Cydonia 24B v2

Overview

Cydonia 24B v2 is a large language model developed by BeaverAI, a project led by TheDrummer. Engineered as a fine-tune of the Mistral "Small" model, Cydonia 24B v2 features 23.6 billion parameters and aims to deliver robust generative capabilities for extended conversational contexts and detailed text generation. The model has garnered attention for its stable performance with long context windows, detailed retention of narrative elements, and adaptability to various chat templates. While further technical and dataset specifics are limited in public sources, user feedback and community reports form the primary basis for evaluating its performance and features.

Technical Specifications

Cydonia 24B v2 is reported to operate with a parameter count of 23.6 billion, leveraging the BF16 tensor format to optimize computational performance. Its architecture is based on a fine-tuned version of the Mistral-Small-24B-Instruct-2501, positioning it within the family of large-scale transformer models oriented toward natural language understanding and generation. Community testing has demonstrated that the model maintains coherence and topicality over extended context windows, with stability observed up to 24,000 tokens and specific user-reported successful dialogue contexts extending to 21,000 tokens without pronounced degradation or repetitive outputs.

Performance and Capabilities

Reports from early users indicate that Cydonia 24B v2 is proficient in maintaining detailed information across lengthy dialogues, demonstrating strength in long-form conversations and complex role-play scenarios. Observers have noted strong retention of narrative continuity, including character, event, and scene details. The model exhibits an extensive vocabulary and demonstrates an ability to fluently handle diverse subject matter. Some users have characterized its generative output as detailed or expressive; however, no systematic content bias or constraint has been identified in available documentation or user discussions.

Training Data and Methodology

While Cydonia 24B v2 is publicly described as a fine-tune of Mistral-Small-24B-Instruct-2501, details on the specific datasets, filtration methods, or training protocols utilized during the fine-tuning process have not been released by the developer. As such, broader information regarding data composition, size, or domain targeting remains unavailable. This omission is consistent with a number of community-released large language models, where development timelines and technical strategies often rely on less formalized, iterative improvement based on feedback and observed behavior.

Use Cases and Applications

Cydonia 24B v2’s ability to manage large contexts with high informational fidelity renders it effective for applications such as interactive storytelling, detailed role-playing sessions, and general text generation tasks. Anecdotal feedback highlights its capacity for handling diverse dialogue scenarios, retaining character voices, and avoiding repetition during multi-thousand-token exchanges. The model supports various popular chat templates, with the Mistral v7 Tekken template recommended for optimal performance. The Metharme template is also supported, albeit with minor compatibility caveats that may necessitate patching.

Limitations and Considerations

Details on explicit model limitations, such as potential biases, factual accuracy, or error characteristics, have not been systematically investigated or published. One practical limitation is that integration with some chat templates (notably Metharme) may require small modifications for seamless operation. Cydonia 24B v2’s licensing details are likewise not specified in public documentation, though the model and its variants are distributed through Hugging Face, with multiple formatting options—including GGUF and EXL2 versions—designed to enhance compatibility across deployment frameworks.

Model Family and Development Context

Cydonia 24B v2 is the successor to earlier iterations such as "Cydonia v1" and occasionally referred to under alternate aliases, including "Cydonia 24B." The model is developed under the BeaverAI project by TheDrummer, noted for a broad professional software background in fields encompassing web development, APIs, and artificial intelligence. While specific release dates and development timelines for each Cydonia version are not publicly documented, ongoing iterative updates signal continued refinement.

Helpful Links

• Cydonia 24B v2 on Hugging Face (GGUF)
• Cydonia 24B v2 iMatrix GGUF (recommended variant)
• Cydonia 24B v2 EXL2 versions
• Community Discussion and Support Channel

For additional developer background and ongoing updates, refer to TheDrummer's project homepage.