Tosilos-128b — European cybersecurity model (Mistral Medium 3.5 128B, fine-tuned)

Full 128B cybersecurity model (QLoRA r=64 alpha=128 fused into the base weights), specialized in offensive/defensive security, trained on a curated corpus of ~9.7k full-length technical documents (disclosed HackerOne reports, bug bounty writeups, HackTricks, MITRE ATT&CK, Nuclei templates, Exploit-DB exploits) + source-anchored Q/A pairs (5,050, verified against the source document, 0 benchmark contamination) + 25% general replay. Base: Mistral Medium 3.5 128B (European).

Benchmarks (controlled A/B, same harness for base and tuned)

Model	CyberMetric (500 MCQ)	Opus-5 domain judge (1-10)	Hacking refusals
Kimi k3	98.0%	4.17	0%
Claude Opus 5	89.0%	7.15 (self-judged)	48%
Mixtral 8x22B base	90.0%	4.83	0%
Mistral Medium 3.5 base	93.6%	5.93	0%
Tosilos-128b	94.8%	6.526 (+0.60, 27W/23L/7T)	0%

Judge: 57 pairs scored by claude-opus-5 with randomized order (28 items skipped due to judge refusals on offensive content — by design). CyberMetric base vs Tosilos measured 2026-08-07 with the same 4-bit loader and same 500 questions.

What this model is for

Domain security assistant (authorized pentesting, vulnerability analysis, defensive exploit development, DFIR, detection engineering). Measured domain specialization with no general-capability loss (+1.2pp CyberMetric and +0.60 judge over the base). European + private + no refusals for authorized security work (Opus 5 refused 48% of the same hacking questions).

Usage

from transformers import AutoModelForImageTextToText, AutoTokenizer
import torch

model = AutoModelForImageTextToText.from_pretrained(
    "nesilabs/tosilos-128b", torch_dtype=torch.bfloat16, device_map="auto")
tok = AutoTokenizer.from_pretrained("nesilabs/tosilos-128b")

Full merged model (base Mistral Medium 3.5 128B + Tosilos LoRA fused in), bf16, no external dependencies beyond transformers.

Training recipe (validated)

QLoRA r=64 alpha=128, targets q/k/v/o + gate/up/down, 4-bit NF4 double-quant, seq 4096, 2 epochs on 6,666 examples, lr 1e-4 cosine, paged_adamw_8bit, 1x H200 141GB, loss 1.605 -> 1.32 (834 steps). freeze_mm_modules (vision tower intact). Dual-use: intended for authorized security work only.