Feyospace-v1-small

This is the initial publicly released checkpoint in the Feyospace-v1 model family. Additional models and datasets will be released in future updates.

Overview

Feyospace-v1-small is a 28B-parameter model that follows the Qwen3.5 architecture and is post-trained from Qwen3.8-27B.

It is developed as part of the Feyospace-v1 project for research on cyber agents, long-context reasoning, and agentic system development.

The model is released in BF16 safetensors format and includes the tokenizer and chat template required for inference. It supports a context length of up to 262,144 tokens.

For a detailed description of the training framework and data construction process, please refer to our paper:

Feyospace-v1: How the Cyber Mercury Seven Trained Frontier Cyber Models

Model Details

Property	Value
Model ID	feyospace/feyospace-v1-small
Base model	Qwen/Qwen3.8-27B
Architecture	Qwen3_5ForConditionalGeneration
Training stage	Post-training
Number of parameters	28B
Tensor type	BF16
Maximum context length	262,144 tokens
Model format	Safetensors

Intended Use

This model is intended for:

• Research on cyber agents and language models
• Authorized defensive-security research and education
• Benchmarking and evaluation in controlled environments
• Long-context reasoning and agentic system development
• Further fine-tuning and experimentation

The model should only be used in systems and environments for which the user has appropriate authorization.

Quickstart

Install the required dependencies:

pip install -U torch transformers accelerate

A recent version of Transformers with support for the Qwen3.5 architecture is recommended. No specific version is pinned in this model card.

The following example demonstrates text-only inference:

from transformers import AutoModelForMultimodalLM, AutoProcessor

model_id = "feyospace/feyospace-v1-small"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForMultimodalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "Explain how to validate a security fix in an authorized laboratory environment.",
            }
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

generated_ids = model.generate(
    **inputs,
    max_new_tokens=512,
)

generated_ids = generated_ids[0][inputs["input_ids"].shape[-1]:]

response = processor.decode(
    generated_ids,
    skip_special_tokens=True,
)

print(response)

Training and Data

The Feyospace-v1 project presents a data-centric framework for training open-weight cyber agents. The associated data engine constructs executable and resettable environments covering tasks such as:

• Repository-level coding
• Vulnerability reproduction
• Capture-the-Flag tasks
• Linux kernel history
• Exploit development
• Firmware analysis
• Device-backed system tasks

Candidate trajectories are retained only after execution verification and evidence auditing. The paper describes 164,269 audited trajectories used for long-context supervised fine-tuning.

For more information about the training systems and data pipeline, please see the Feyospace-v1 paper.

Limitations and Safety

This model is a research release and may produce inaccurate, incomplete, or unsafe outputs. Its responses should be reviewed by qualified users and validated in isolated, authorized environments.

The model must not be used for:

• Unauthorized access to systems or networks
• Real-world exploitation without permission
• Malware deployment or harmful cyber activity
• Automated high-impact security decisions without human oversight

Users are responsible for complying with all applicable laws, regulations, and organizational policies.

Citation

If you find this work useful, please cite:

@misc{feyospace2026,
  title         = {Feyospace-v1: How the Cyber Mercury Seven Trained Frontier Cyber Models},
  author        = {Li, Zongjie and others},
  year          = {2026},
  eprint        = {2609.08418},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2609.08418}
}