Alexandria-based TokenAI has released Horus Cyber Nano 1.0, a 16-billion-parameter multimodal model designed for cybersecurity, secure coding and defensive technical work.
The model’s compact open weights were released on 8 September 2026 through Hugging Face, alongside GGUF variants for local deployment. It can process text, images and video, supporting use cases including secure code review, visual inspection and video understanding.
Horus Cyber Nano uses a mixture-of-experts architecture with 64 routed experts, two shared experts and six active per token. It has a maximum context length of 131,072 tokens and can process images of up to 3.2 million pixels.
The release followed TokenAI’s decision to pause the full-size Horus Cyber model to meet additional open-weight safety requirements. According to the company, Horus Cyber Nano outperformed OpenAI’s GPT-4o on some vision and reasoning tests and came close to Qwen2.5-VL-72B on several multimodal tasks. TokenAI’s published benchmarks include scores of 84.4 on MMBench-EN-v1.1, 91.8 on MATH, 82.0 on MMLU and 91.4 on ScreenSpot-V2, while noting that testing conditions vary.
TokenAI said it is developing a further Nano release focused on stronger capabilities, higher performance and lower computational costs. Its Horus Cyber Pro models remain closed source, with open weights planned in the coming months pending approval from Egypt’s National AI Council.
Source: Middle East AI News
