Horus Cyber Nano 1.0
Official model page: Horus Cyber Nano 1.0
Horus Cyber Nano 1.0 is TokenAI's compact open-weights multimodal cybersecurity model for secure code review, vulnerability analysis, authorized red-team reasoning, terminal workflows, tool-use reasoning, and image-grounded technical understanding.
Overview
Horus Cyber Nano 1.0 is built for practical defensive security work. The model accepts text and images as input and produces text output, making it useful for code review, terminal-centric investigation, screenshot analysis, and long-context technical reasoning.
Developed by TokenAI and founded by Assem Sabry, Horus Cyber Nano 1.0 is part of the Horus model family and is positioned as a specialized model for security-heavy engineering workflows rather than a general-purpose assistant.
Model Links
- Main model: tokenaii/Horus-Cyber-Nano-1.0
- GGUF repo: tokenaii/Horus-Cyber-Nano-1.0-GGUF
- Technical page: tokenai.llc/models/horus-cyber-nano1.0
What It Can Do
1. Secure Code Review
Analyze source code for security weaknesses, explain root causes, map issues to secure engineering patterns, and suggest safer remediation paths.
2. Red Team Task Reasoning
Break down authorized security-testing tasks into structured steps, hypotheses, and evidence-oriented analysis suitable for controlled environments.
3. Terminal and Tool Workflows
Understand command-line tasks, propose useful commands, interpret outputs, and support multi-step technical workflows involving tools and system state.
4. Cybersecurity Vision Understanding
Read screenshots, code images, and terminal captures to extract technical details and convert them into actionable text analysis.
Technical Details
The following specifications are transcribed from the official Horus Cyber Nano 1.0 model page.
| Item | Official value |
|---|---|
| Architecture | Native multimodal vision-language model with a Mixture-of-Experts causal language decoder |
| Release | September 2026 |
| License | TokenAI Custom License |
| Parameter count | 16B parameters |
| Tensor types | BF16 and F32 |
| Input and output | Text, images, and video in; text out |
| Maximum image resolution | 3.2 million total pixels per image |
| Transformer layers | 27 |
| Hidden size | 2,048 |
| Attention heads | 16 |
| KV heads | 16 |
| MoE | Yes |
| Experts per MoE layer | 64 routed experts plus 2 shared experts |
| Expert parallel size | 1 |
| Active experts per token | 6 routed experts |
| MoE layer frequency | Every layer |
| Activation | SiLU |
| RoPE theta | 800,000 |
| Vision encoder layers | 27 |
| Vision hidden size | 1,152 |
| Vision attention heads | 16 |
| Vision patch size | 14 |
Official Benchmarks
Results and comparison values below are taken from the official model page. They are presented as published; evaluation protocols may differ across providers unless a protocol is explicitly stated.
MMLU
| Model | Score |
|---|---|
| Horus Cyber Nano 1.0 | 82.0% |
| OpenAI o1 | 91.8% |
| GPT-4.5 | 90.8% |
| GPT-4.1 | 90.2% |
| Grok-2 | 87.5% |
| DeepSeek V3.2 Thinking | 85.0% |
GPQA Diamond
| Model | Score |
|---|---|
| Horus Cyber Nano 1.0 | 45.2% |
| Mistral 3.1 24B | 45.96% |
| GPT-4.1 | 66.3% |
| gpt-oss-120b | 67.1% |
| DeepSeek-V3.2 | 82.4% |
Terminal-Bench 2.0
| Model | Score |
|---|---|
| Horus Cyber Nano 1.0 | 46.2% |
| GPT-5.5 | 82.7% |
| GPT-5.4 | 75.1% |
| Gemini 3.1 Pro Thinking | 68.5% |
| GPT-5.2 Thinking | 62.2% |
| Claude Sonnet 4.6 | 59.1% |
AIME 2025
| Model | Score |
|---|---|
| Horus Cyber Nano 1.0 | 76.4% |
| GPT-5.2 Thinking | 100.0% |
| Claude Sonnet 4.6 | 95.6% |
| GPT-5 | 94.6% |
| Gemini 3 Flash | 95.2% |
| DeepSeek-R1-0528 | 87.5% |
SWE-Bench Verified
| Model | Score |
|---|---|
| Horus Cyber Nano 1.0 | 61.7% |
| Gemini 3.1 Pro Thinking | 80.6% |
| GPT-5.2 Thinking | 80.0% |
| Claude Sonnet 4.6 | 79.6% |
| GPT-5 | 74.9% |
| Claude Opus 4 | 72.5% |
Multimodal benchmarks
| Benchmark | Horus Cyber Nano 1.0 | GPT-4o | Qwen2.5-VL-72B | Qwen2.5-VL-32B | Gemma 3 27B |
|---|---|---|---|---|---|
| MATH-Vision | 56.9% | 30.4% | 38.1% | 38.4% | 35.4% |
| MathVista MINI | 80.1% | 63.8% | 74.8% | 74.7% | 59.8% |
| MMBench-EN-v1.1 | 84.4% | 83.1% | 88.3% | - | 78.9% |
| MMStar | 70.4% | 64.7% | 70.8% | 69.5% | 63.1% |
| VideoMMMU | 65.2% | 61.2% | 60.2% | - | 61.8% |
GGUF Variants
GGUF variants are available in the companion repository: tokenaii/Horus-Cyber-Nano-1.0-GGUF
| Variant | Approx. combined RAM + VRAM | Download |
|---|---|---|
| F16 | 59.71 GiB | GGUF repo |
| Q8_0 | 32.18 GiB | GGUF repo |
| Q6_K | 27.24 GiB | GGUF repo |
| Q5_K_M | 23.11 GiB | GGUF repo |
| Q4_K_M | 20.84 GiB | GGUF repo |
Note: Combined RAM and VRAM usage was measured approximately for each variant under a fixed local test configuration. Actual requirements may vary with context length, generation length, batching, runtime, and hardware.
Quick Start
Using NeuralNode
import neuralnode as nn
model = nn.Provider.horus(
model="horus-cyber-nano-1.0",
device="cuda"
)
agent = nn.Agent(
role="Security Engineer",
model=model,
system_prompt="You analyze code, terminal output, and screenshots for defensive cybersecurity work."
)
result = agent.execute(
task="Review this authentication flow for insecure patterns and propose a safer patch."
)
print(result)
You will also be able to use Horus Cyber Nano 1.0 through Ollama and LM Studio, and fine-tune it with Unsloth. Direct runtime and fine-tuning links will be added with the public release materials. Microsoft Foundry support is planned.
Intended Use
- Secure code review
- Vulnerability triage and remediation guidance
- Terminal-centric engineering workflows
- Tool-augmented reasoning for defensive security analysis
- Screenshot and code-image understanding
Safety Note
Horus Cyber Nano 1.0 is intended for authorized security analysis and defensive workflows. Public benchmarking, documentation, and deployment guidance should be interpreted within that scope.
Note on Identity Responses
As with any open-weight model, occasional hallucinations or inconsistent answers about the model's name and identity may occur. This can result from training techniques, decoding settings, and differences among training-data sources; users should verify identity claims against the official model page and release artifacts.
Contact & Community
- Website: tokenai.llc
- Hugging Face: huggingface.co/tokenaii
- LinkedIn: linkedin.com/company/tokenaii
- Facebook: facebook.com/tokenaii
- X: x.com/tokenaia
- Contact: info@tokenai.llc
Citation
@misc{tokenai_horus_cyber_nano_1_0_2026,
title = {Horus Cyber Nano 1.0},
author = {TokenAI},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/tokenaii/Horus-Cyber-Nano-1.0}},
note = {Multimodal MoE model for secure coding, terminal workflows, and cybersecurity reasoning}
}
License
Horus Cyber Nano 1.0 is released under the TokenAI Custom License. Final license text and release usage terms should be consulted alongside the public model release materials.