FROSTWOLF LABS · OPEN WEIGHTS

FROST CYBER MINI 1.0: A 4B MODEL THAT SPEAKS OFFENSE.

Frost Cyber Mini 1.0 is our compact copilot for red teams, OSINT and penetration testing. It explains CVEs and ATT&CK tradecraft, drafts proofs of concept, and answers the questions general-purpose models refuse. It runs on your laptop, so nothing leaves the machine.

Specification

  • PARAMETERS: 4.2B · BF16
  • BASE: Qwen3.5-4B (abliterated)
  • ARCHITECTURE: Hybrid · Gated-DeltaNet + full attention
  • CONTEXT: 262k native
  • RUNS ON: A 16 GB laptop
  • LICENSE: CC-BY-NC-SA-4.0 · gated

Benchmarks

Multiple-choice security knowledge, scored locally. The card's own reading: cyber knowledge in the range of much larger models.

  • CyberMetric-500: 87.8% (439 / 500), Cybersecurity knowledge
  • CTIBench · cti-mcq: 70.4% (1,759 / 2,500), Threat intelligence and ATT&CK

llama.cpp · Q8_0 GGUF · greedy decoding · letter-parsed multiple choice. These measure knowledge and reasoning, not autonomous exploitation.

Downloads

  • Safetensors: BF16 · full precision · Transformers, vLLM, SGLang
  • GGUF: Q4_K_M to Q8_0 · llama.cpp, Ollama, LM Studio
  • MLX 4-bit: 2.2 GB · fits in 16 GB · the Apple-silicon default
  • MLX 8-bit: 4.2 GB · near-lossless · 32 GB and up

Responsible use

This model is abliterated: its safety filtering is substantially reduced, and it will write exploit code and attacker tradecraft on request. That is what makes it useful to a red team, and why access is gated.

  • Authorized work only: Systems you own or are explicitly permitted to test. Unauthorized access, real-world attacks, fraud and harassment are prohibited.
  • Non-commercial: Released under CC-BY-NC-SA-4.0 because some training data is non-commercial. Commercial use is not permitted under this release.
  • Verify everything: A 4B model can be confidently wrong. Check CVE, CVSS and CWE details against primary sources, and treat exploit output as an untrusted draft.