AuroraSystem/Ficus-0.1-Agent-12B

Ficus 0.1 Agent 12B: A Cybersecurity-Focused LLM

Ficus 0.1 Agent 12B, developed by Aurora System, is a 12 billion parameter open-weight language model based on google/gemma-4-12B-it. It is specifically fine-tuned for cybersecurity tasks and bilingual (Russian/English) contexts, with an impressive 32768 token context length. While it includes experimental agentic capabilities and a built-in thinking mode, developers explicitly state that Ficus 0.1 is an early test model and not production-ready for agentic pipelines, as its tool-calling functionality is currently unreliable and fabricates results.

Key Capabilities

• Cybersecurity Expertise: Strong performance in vulnerability analysis, source code auditing, binary exploitation (pwn/rev), applied cryptography, CVE breakdowns, and penetration testing methodology.
• Web Security Auditing: Proficient in analyzing OWASP Top 10 vulnerabilities such as SQL injection, JWT authentication bypass, CSRF, SSRF, and IDOR/BOLA.
• Reverse Engineering: Capable of analyzing x86_64/ARM assembly listings, recovering C pseudocode, and analyzing custom virtual machines.
• Reasoning: Features a built-in thinking mode (<|channel>thought) for complex reasoning, though caution is advised due to potential degenerate loops.
• Multimodal Input (Inherited): Supports image, audio, and video input, inherited from its base model, though fine-tuning was text-only and multimodal behavior was not validated.

Performance Highlights

Ficus 0.1 demonstrates significant improvements in cybersecurity benchmarks compared to its base model:

• SecEval: 73.00% (vs. 62.80% for base Gemma 4 12B)
• SecQA v1: 100.00% (matches base)
• SecQA v2: 97.00% (matches base)
• CyberMetric-500: 92.20% (vs. 89.80% for base)

It also shows gains in general knowledge benchmarks like MMLU (+7.2 points) and MMLU-Pro (+23.2 points) when thinking mode is enabled. However, it exhibits regressions in code generation (LiveCodeBench pass@1: 34.8% vs. 54.5% for base) and multilingual performance (ruMMLU -4.4, MMMLU-ZH -7.2).

Should I use this for my use case?

Use Ficus 0.1 Agent 12B as a text assistant for:

• Cybersecurity research and education: Ideal for analyzing vulnerabilities, auditing code, and understanding penetration testing methodologies.
• CTF competitions: Can assist with tasks related to pwn/rev, applied cryptography, and general security problem-solving.
• Starting point for further fine-tuning: Its specialized dataset makes it a strong foundation for developing more advanced cybersecurity-focused models.

Do NOT use this model for:

• Production agentic pipelines: The model's agentic capabilities are experimental and known to fabricate tool calls and results. Connecting it to real systems, filesystems, or networks is strongly discouraged.
• Code generation: It shows a significant regression in code generation performance compared to its base model.
• General-purpose multilingual tasks: There is a noted drop in performance for Russian and Chinese language benchmarks.