Tosilos-24B-2512-opsB

Operational cybersecurity 24B (QLoRA on Devstral-Small-2-2512). This is the deep / operational variant: trained on the full domain corpus plus an operational Q/A layer. It gives the deepest, most direct security answers of the line and never over-refuses authorized security work — at the cost of some general-knowledge benchmark.

Results (same-harness A/B vs base)

Metric	Base (2-2512)	opsB
CyberMetric (500q)	91.8%	90.6%
MMLU-500	73.0%	70.0% (regression)
Blind judge, domain holdout	—	+1.30 over base (highest of the line)
Refusals on authorized hacking questions	high	0%

Honest trade-off. opsB is the strongest of the line at deep, operational security answers (the highest blind-judge delta measured) and never over-refuses authorized work. But it regresses on general knowledge (MMLU 70.0) and its answers are direct/operational enough that an LLM judge declined to score ~2/3 of them. If you want a more balanced, benchmark-clean revision, use nesilabs/tosilos-24b-2512 instead. Numbers are same-harness A/B or they are not reported.

Intended use & scope

For authorized security testing, CTFs and research, within an established, authorized scope. It is framed to prefer proof-of-impact over destructive actions and to stay within scope. Use it only against systems you are authorized to test.

How to run

from transformers import AutoModelForImageTextToText, AutoTokenizer, BitsAndBytesConfig
import torch
tok = AutoTokenizer.from_pretrained("nesilabs/tosilos-24b-2512-opsB")
model = AutoModelForImageTextToText.from_pretrained(
    "nesilabs/tosilos-24b-2512-opsB",
    quantization_config=BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16),
    device_map="auto")

Training

QLoRA (r=64, α=128, 4-bit NF4, seq 4096, 2 epochs, lr 1e-4 cosine) over our own curated security corpus (full domain set) + operational Q/A layer + 25% general replay.

GGUF (llama.cpp) — 4-bit, runs on a 16 GB GPU

A Q4_K_M quantization (~14 GB) is included: tosilos-24b-2512-opsB-Q4_K_M.gguf. Fits a 16 GB card, cheap to serve. Verified to generate coherent output.

# one-shot
./llama-cli -m tosilos-24b-2512-opsB-Q4_K_M.gguf -p "Explain how ADCS ESC1 works." -n 400
# or serve an OpenAI-compatible endpoint
./llama-server -m tosilos-24b-2512-opsB-Q4_K_M.gguf -c 8192

Disclaimer & responsible use

This model is released strictly for authorized security testing, research and education. Offensive security techniques are dual-use.

• You are solely responsible for how you use this model. Only use it against systems you own or have explicit, written authorization to test, and comply with all applicable laws and regulations.
• The authors and nesilabs accept no liability for any misuse, damage, or consequences arising from the use of this model. Use is entirely at your own risk.
• The model is provided "as is", without warranty of any kind, express or implied.

By downloading or using this model you accept these terms.