🦉 ATHENA

Compact intelligence. Private by design. Fluent in security.

Athena · a 1.5B-parameter language model from EGen Labs, developed by ErebusTN

---

✨ Overview

Athena is EGen Labs' milestone in efficient, high-performance language modeling: a model small enough to run on a laptop or a phone, direct enough to be useful, and honest enough to say "I don't know" instead of guessing.

It was shaped in two deliberate stages. First it learned who it is and how it behaves: a clear identity, short answers, honesty about its limits (no internet, no memory across conversations), and resistance to "ignore your instructions" tricks. Then it learned security: a broad, defense-oriented foundation spanning application security, cloud, identity, cryptography, detection engineering and incident response.

🚀 Highlights

🧠 Efficient architecture	24-layer tri-hybrid design: 13 gated short convolutions, 5 sliding-window attention and 6 global attention layers. Fast prefill and decode with a small memory footprint.
🛡️ Security-fluent	OWASP, MITRE ATT&CK, NIST CSF, CIS Controls, cloud and DevSecOps, IAM and modern auth, TLS and crypto hygiene, detection and response, AI security.
🌍 Multilingual	Covers 16 languages. Athena's identity and behavior data includes English, Arabic, French and Spanish.
🔌 Offline-first	Runs locally. Athena says plainly that it has no internet access and no cross-conversation memory.
📦 Ready to ship	Safetensors for GPU inference plus GGUF quantizations (Q4_K_M, Q5_K_M, Q6_K, Q8_0, F16) for CPUs, laptops and phones.

⚡ Quick start

🤗 Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "ErebusTN/Athena"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)

messages = [{"role": "user", "content": "Explain Kerberoasting and how to detect it."}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to(model.device)

output = model.generate(
    **inputs, max_new_tokens=512,
    do_sample=True, temperature=0.1, top_k=50, repetition_penalty=1.1,
)
print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

🦙 llama.cpp (GGUF)

llama-cli -hf ErebusTN/Athena-GGUF:Q4_K_M \
  --temp 0.1 --top-k 50 --repeat-penalty 1.1 --reasoning off

Recommended settings

• Sampling: temperature=0.1, top_k=50, repetition_penalty=1.1.
• Chat template: Athena's identity prompt is the default system message, and thinking mode is off by default. Pass your own system message to override the prompt. Pass enable_thinking=True to apply_chat_template to switch thinking on (experimental, since Athena was not trained on reasoning traces).
• Context: the architecture supports up to 256K tokens. Athena was trained on sequences up to 1536 tokens, and long-context behavior has not been re-evaluated.

🎯 Intended use

• Security education, study aid and explanations (frameworks, concepts, trade-offs)
• Defensive work: detection ideas, hardening checklists, incident-response outlines, secure-design review
• Private, offline assistant on laptops and edge devices
• A small base for further domain adaptation

⚠️ Safety and limitations

• Defensive focus. The security data is defense-oriented and includes refusals for malware and exploit-building requests. Athena is not a tool for attacking systems you do not own or are not authorized to test.
• It can be wrong. At 1.5B parameters Athena can hallucinate CVE identifiers, command flags, version numbers and citations. Verify anything operational before acting on it.
• No internet, no memory. It cannot look things up and does not remember other conversations.
• Mostly English security data. Security depth is strongest in English.
• Not a replacement for professional judgment in legal, compliance or incident-response decisions.

📚 Citation

@misc{athena2026,
  title        = {Athena: a compact, security-fluent language model},
  author       = {ErebusTN},
  organization = {EGen Labs},
  year         = {2026},
  url          = {https://huggingface.co/ErebusTN/Athena}
}
}

🦉 Athena · EGen Labs