feyospace/feyospace-v1-small
Feyospace/feyospace-v1-small is a 28-billion parameter model based on the Qwen3.5 architecture, post-trained from Qwen3.8-27B. Developed for research on cyber agents, it supports an extensive context length of up to 262,144 tokens. This model is specifically designed for long-context reasoning and agentic system development, particularly in authorized defensive-security research and education. Its training framework focuses on constructing executable and resettable environments for cyber tasks like vulnerability reproduction and exploit development.
Loading preview...
Overview
Feyospace-v1-small is a 28-billion parameter model built on the Qwen3.5 architecture, post-trained from the Qwen3.8-27B base model. It is the initial public release in the Feyospace-v1 family, developed for advanced research in cyber agents, long-context reasoning, and agentic system development. The model supports an exceptionally long context length of up to 262,144 tokens and is released in BF16 safetensors format.
Key Capabilities & Features
- Cyber Agent Research: Specifically designed to facilitate research into cyber agents, including their development and application.
- Extended Context Window: Features a maximum context length of 262,144 tokens, enabling deep analysis and reasoning over large amounts of information.
- Specialized Training: Utilizes a data-centric framework for training, involving an engine that constructs executable and resettable environments for various cyber tasks. This includes repository-level coding, vulnerability reproduction, Capture-the-Flag tasks, exploit development, and firmware analysis.
- Audited Trajectories: Training data includes 164,269 audited trajectories, ensuring verified execution and evidence for long-context supervised fine-tuning.
Intended Use Cases
- Defensive Security Research: Ideal for authorized defensive-security research and educational purposes.
- Agentic System Development: Suitable for developing and experimenting with agentic systems that require sophisticated reasoning capabilities.
- Benchmarking & Evaluation: Can be used for benchmarking and evaluation in controlled environments.
- Further Fine-tuning: Provides a strong base for further fine-tuning and experimentation in specialized cyber domains.
Limitations
As a research release, the model may produce inaccurate, incomplete, or unsafe outputs. It is crucial that responses are reviewed by qualified users and validated in isolated, authorized environments. The model is explicitly not to be used for unauthorized access, real-world exploitation without permission, malware deployment, or automated high-impact security decisions without human oversight.