---
title: "Blackfrost-AI's CYBER-FROST Brings a 180B Security Model to Local Machines"
subtitle: "A 180B security-focused Mixture of Experts model is now available in eight local-friendly GGUF quants, with speeds up to 2.4 tokens per second on a single GPU."
url: "https://alphasignal.ai/news/blackfrost-ai-s-cyber-frost-brings-a-180b-security-model-to-local-machines"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-09-27T22:57:03.000Z"
updated: "2026-09-30T22:04:53.709Z"
type: "Model"
topics:
  - "Llms"
  - "Gpus"
subtopics:
  - "Mixture Of Experts"
  - "Small Models"
  - "Quantization"
upvotes: 114
source_url: "https://huggingface.co/freakyskittle/CYBER-FROST-3.8-GGUF"
comments: 0
truncated: true
---

# Blackfrost-AI's CYBER-FROST Brings a 180B Security Model to Local Machines

> A 180B security-focused Mixture of Experts model is now available in eight local-friendly GGUF quants, with speeds up to 2.4 tokens per second on a single GPU.

# CYBER-FROST brings a 180B security model to GGUF

[CYBER-FROST-3.8-GGUF](https://huggingface.co/freakyskittle/CYBER-FROST-3.8-GGUF) packages Blackfrost-AI’s Qwen-based cybersecurity fine-tune as local GGUF files for llama.cpp. The release offers eight quantized variants, with reported file sizes from 80 GB to 123 GB, so developers can balance model fidelity, storage, memory pressure, and decode speed on a high-memory workstation. Security teams gain a way to keep sensitive engagement data off hosted APIs, although benchmark coverage and long-context behavior remain limited.

## A 180B model that activates 6B

The underlying BF16 checkpoint is a Mixture of Experts model based on Qwen3.8-Flash-Next. It contains about 180 billion parameters but routes each token through 10 of 512 experts plus a shared expert, activating roughly 6 billion parameters per token. That routing reduces computation per forward pass, while the full expert pool still requires substantial storage and memory bandwidth.

__Core architecture__
| Component              | Specification                                                     |
| ---------------------- | ----------------------------------------------------------------- |
| Text stack             | 48 blocks with hybrid linear and full attention                   |
| Full-attention cadence | Every fourth block                                                |
| Hidden size            | 2,560                                                             |
| Attention heads        | 24 query heads and 2 key-value heads                              |
| Expert routing         | 512 routed experts, 10 selected per token, plus one shared expert |
| Parameters             | About 180B total and 6B active per token                          |
| Configured context     | Up to 262,144 tokens                                              |

## Security refusals are the tuning target

Security prompts frequently combine terms associated with both legitimate testing and malicious activity. General-purpose assistants may refuse requests involving exploit validation, malware analysis, incident response, or detection engineering even when the operator is working within an approved scope. Blackfrost’s [BF16 checkpoint](https://huggingface.co/Blackfrost-AI/CYBER-FROST-3.8-BF16) was fine-tuned to reduce those false refusals for authorized offensive and defensive workflows.

The model card lists coverage across several security domains:

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [CYBER-FROST-3.8-GGUF model card](https://huggingface.co/freakyskittle/CYBER-FROST-3.8-GGUF)
- [CYBER-FROST-3.8-BF16 base checkpoint](https://huggingface.co/Blackfrost-AI/CYBER-FROST-3.8-BF16)
- [Qwen3.8-Flash-Next foundation model](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)