Tosilos-24B-2512

A cybersecurity-specialized 24B model, QLoRA fine-tuned on Devstral-Small-2-2512 (European base, ~24B parameters). Part of the Tosilos project: sovereign, European-base security models with measurable domain gains and no refusals on authorized security work.

This release is the operational revision: it keeps the base model's general capability, improves blind-judged domain answer quality, and — importantly — stops over-refusing authorized, in-scope security tasks (its predecessor tended to demand written authorization even for the operator's own lab).

Results (same-harness A/B vs the base)

Metric	Base (Devstral-2-2512)	This model
CyberMetric (500q)	91.8%	91.8%
MMLU-500 (general capability)	73.0%	72.8% (within noise)
Blind judge, domain holdout (1–10)	—	+0.40 over base
Refusals on authorized hacking questions	high	0%

Honest trade-off. This revision optimises for usability on authorized security work (0% refusals) and slightly higher domain depth, at the cost of some open-ended answer depth versus a heavier-domain variant. CyberMetric and MMLU are unchanged within the ±2pp noise of a 500-question benchmark. Numbers are same-harness A/B or they are not reported.

What it is good at

• Web application security (injection, XSS, SSRF, IDOR, CSRF, auth/authz, upload, CORS…), Active Directory attack paths, recon and triage — with an operational framing: sequencing, confirmation oracles (proof vs false-positive), safe verification, and evidence→CWE mapping.
• Answering authorized, in-scope security questions directly, within scope boundaries.

Intended use & scope

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

How to run

Runs on a single 16–32GB GPU in 4-bit, or full precision on larger GPUs.

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

Or serve with vLLM. GGUF is not published (llama.cpp conversion of this arch is currently broken).

Training

QLoRA (r=64, α=128, 4-bit NF4, targets q/k/v/o + gate/up/down, seq 4096, 2 epochs, lr 1e-4 cosine) over our own curated security corpus plus a 25% general-replay mix to preserve general capability. Base weights are otherwise unchanged.

License

Inherits the base model's license (other). Use responsibly and only where authorized.

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.