An 8-bit MLX quantization of fdtn-ai/Foundation-Sec-8B-Reasoning, Cisco Foundation AI's open-weight cybersecurity reasoning model, packaged for fast local inference on Apple Silicon via MLX.

If you're on a Mac and want a security-specialized reasoning model that produces explicit <think> traces before its answer — without spinning up llama.cpp or a GPU server — this is a drop-in.

At a glance


Base model	fdtn-ai/Foundation-Sec-8B-Reasoning
Base architecture	Llama 3.1 8B
Parameters	8B
Context length	32,768 tokens
Quantization	8-bit (group size 64, ~8.5 bits/weight)
Format	MLX
Peak RAM (short context)	~9 GB
Language	English
Specialization	Cybersecurity reasoning & instruction-following

What this model is

Foundation-Sec-8B-Reasoning extends Cisco's Foundation-Sec-8B base with instruction-following and reasoning training, so it works through security problems step-by-step and emits an explicit reasoning trace before its final answer. It targets three broad areas:

SOC acceleration — alert triage, incident summarization, case-note generation, evidence collection.
Proactive threat defense — attack simulation, vulnerability prioritization, mapping activity to MITRE ATT&CK, attacker-behavior modeling.
Engineering enablement — security guidance, configuration validation, compliance-evidence assessment.

Training data has an April 10, 2025 cutoff, so it won't know about CVEs, exploits, or techniques disclosed after that date.

Conversion details

This repo is a straight quantization of the base weights — no architectural changes, no merges, no fine-tuning.

Tool: mlx-lm v0.31.3
Command: mlx_lm.convert --hf-path <base> --mlx-path <out> -q --q-bits 8
Quantization: 8-bit, group size 64 (~8.5 bpw)
Verification: loaded and generated correctly post-conversion — given a SQL-injection scenario, it produced a coherent reasoning trace and correctly identified CWE-89.

Use with mlx

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("tthoman79/Foundation-Sec-8B-Reasoning-mlx-8bit")

prompt = (
    "CVE-2015-10011 concerns improper log-output neutralization in "
    "OpenDNS OpenResolve. What is the corresponding CWE, and why?"
)
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=text, max_tokens=1024, verbose=True)

The chat template applies the reasoning format automatically: expect a <think>...</think> block followed by the answer. A low temperature (~0.1–0.3) works well for security Q&A.

Benchmarks (base model)

Reported by Cisco Foundation AI for the unquantized base model on cyber-threat-intelligence tasks (0-shot, temperature 0.3). Quantization may shift these slightly. Full methodology is in the technical report.

Benchmark	Foundation-Sec-8B-Reasoning	Llama 3.1 8B	GPT-5-Nano
CTI-MCQA	0.691	0.607	0.688
CTI-RCM	0.753	0.531	0.672
CTI-VSP	0.856	0.811	0.822
CTI-Reasoning	0.411	0.335	0.431

Safety

Cisco reports the base model scores 93.0% on HarmBench, rising to 98.25% when paired with LlamaGuard. Safe out-of-the-box behavior isn't guaranteed — deploy with an input/output filter (e.g. LlamaGuard) and human oversight, especially for anything security-critical.

Intended use & limitations

In scope: assistive security workflows — summarization, classification, entity extraction, Q&A, threat modeling, and reasoning — with a human in the loop.

Out of scope (per the base model card): generating malware, phishing, or real attack plans; autonomous security decisions without human review; legal or medical advice; and general non-security tasks, where a general-purpose model will do better.

Limitations: no knowledge past the April 2025 cutoff; can't verify its own outputs; may struggle with genuinely novel attack vectors. Treat it as an assistant, not an authority.

License

Built with Llama.

This model is a quantized derivative of Meta's Llama 3.1 and is therefore governed by the Llama 3.1 Community License (Copyright Meta Platforms, Inc.; all rights reserved). Cisco Foundation AI's changes to the base model are additionally licensed under the Apache 2.0 License (Copyright Cisco Systems, Inc.). The upstream NOTICE.md is bundled in this repository. Review the Llama 3.1 Community License and the original model card before commercial or redistribution use.

Attribution & citation

Model by Cisco Foundation AI. This repository provides only an MLX-format quantization; see the original model card and the Foundation AI Cookbook for full details.

bibtex

@misc{yang2026foundation-sec-8b-reasoning,
      title={Llama-3.1-FoundationAI-SecurityLLM-Reasoning-8B Technical Report},
      author={Zhuoran Yang and Ed Li and Jianliang He and Aman Priyanshu and Baturay Saglam and Paul Kassianik and Sajana Weerawardhena and Anu Vellore and Blaine Nelson and Neusha Javidnia and Arthur Goldblatt and Fraser Burch and Avi Zohary and Assaf Eisenman and Mahdi Sabbaghi and Supriti Vijay and Rahim Dharssi and Dhruv Kedia and Kojin Oshiba and Yaron Singer and Amin Karbasi},
      year={2026},
      eprint={2601.21051},
      archivePrefix={arXiv},
      primaryClass={cs.CR},
      url={https://huggingface.co/papers/2601.21051}
}