Tosilos-24b — home-runnable European cybersecurity model (Mistral Devstral 24B, fine-tuned)

Full 24B cybersecurity model (QLoRA r=64 alpha=128 fused into Devstral-Small-2505), trained on the Tosilos curated corpus (17k domain examples + 25% general replay). Designed to run locally on a single 32GB-class GPU (RTX 5090) in 4-bit.

Benchmarks (same harness, 500 questions each; judge: Opus 5, blind, randomized)

Metric	Devstral 24B base	Tosilos-24b
CyberMetric	91.6%	92.8% (+1.2pp)
MMLU (general knowledge)	77.2%	77.4% (no forgetting)
Domain judge (Opus 5, 1-10)	5.49	5.68 (+0.19, 26W/23L/8T)

Tosilos-24b improves over its base on all three axes — modest but consistent. Recipe note: the 2-epoch run (834 steps over 6,666 examples) beat the 1-epoch XL run (17k examples), which had matched the base — epoch count mattered more than dataset size.

Usage (home server, 1× 32GB GPU)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "nesilabs/tosilos-24b", load_in_4bit=True, device_map="auto")
tok = AutoTokenizer.from_pretrained("nesilabs/tosilos-24b")

Or vLLM tensor-parallel across 2 GPUs: vllm serve nesilabs/tosilos-24b -tp 2.

Files

• model-*.safetensors — full merged model, bf16 (48GB). A GGUF Q4 build was evaluated but not published: current llama.cpp conversions of this architecture (tekken tokenizer, rope_theta=1e9) degrade output. Prefer the safetensors path until llama.cpp support matures.

Training recipe

QLoRA r=64 alpha=128, targets q/k/v/o + gate/up/down, 4-bit NF4, seq 4096, 2 epochs on 6,666 examples, lr 1e-4 cosine, 1×H200, 834 steps. Base: mistralai/Devstral-Small-2505.

Dual-use: intended for authorized security work only.

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.