EUAIAct-Qwen2.5-0.5B-Edge
A compact, offline-first Language Model specialized in EU AI Act & GDPR
Runs on mobile browsers · laptops · edge servers — no cloud, no API, no data transfer
Overview
EUAIAct-Qwen2.5-0.5B-Edge is a 494M parameter Small Language Model fine-tuned on the EU AI Act (Regulation EU 2024/1689) and GDPR compliance corpus.
It is designed to run entirely on-device — in a mobile browser via WebGPU, on Apple Silicon via MLX, or on any machine via GGUF. No internet connection required after download.
Advisory use only — not a substitute for qualified legal counsel.
✦ Key Features
- Offline-first — works without internet once downloaded
- Cross-platform — Windows · macOS · Linux · Mobile · Browser
- Mobile-ready — runs in browser via Transformers.js + WebGPU
- Multilingual — French 🇫🇷 and English 🇬🇧
- Lightweight — 350 MB (GGUF) to 1 GB (MLX full precision)
- Open weights — Apache 2.0, fully auditable
Deployment formats
| Format | Size | Runtime | Best for |
|---|---|---|---|
| ONNX INT8 | ~400 MB | Transformers.js | Browser · Mobile · WebGPU — Windows · macOS · Linux |
| GGUF Q4_K_M | ~350 MB | llama.cpp | Windows · macOS · Linux · Edge servers |
| MLX BF16 | ~1.0 GB | mlx_lm | Apple Silicon only (M1/M2/M3/M4) |
Quickstart
🌐 Browser & Mobile — Transformers.js + WebGPU
import { pipeline } from '@xenova/transformers';
const ai = await pipeline(
'text-generation',
'sallani/EUAIAct-Qwen2.5-0.5B-Edge',
{ device: 'webgpu' }
);
const response = await ai(
"What are the obligations for high-risk AI system providers under Article 16?",
{ max_new_tokens: 400 }
);
console.log(response[0].generated_text);
🍎 Apple Silicon — MLX
from mlx_lm import load, generate
model, tokenizer = load("sallani/EUAIAct-Qwen2.5-0.5B-Edge")
messages = [
{"role": "system", "content": "Tu es un expert EU AI Act et RGPD."},
{"role": "user", "content": "Qu'est-ce qu'un système IA à haut risque selon l'Article 6 ?"}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512))
🖥️ llama.cpp — Windows · macOS · Linux
llama-cli \
-m euaiact-qwen2.5-0.5b-q4_k_m.gguf \
--chat-template qwen \
-p "Explain the EU AI Act conformity assessment procedure." \
-n 512
Model details
| Property | Value |
|---|---|
| Base model | Qwen/Qwen2.5-0.5B-Instruct (Apache 2.0) |
| Fine-tuning method | LoRA — r=8, α=16, 8 layers |
| Training hardware | Apple M4 Max · MLX |
| Parameters | 494M |
| Training data | 300 curated Q&A pairs — EU AI Act · GDPR |
| Languages | French 🇫🇷 · English 🇬🇧 |
| Context length | 1 024 tokens |
| License | Apache 2.0 |
Training corpus
Covers the full regulatory stack for AI compliance in the EU:
| Regulation | Coverage |
|---|---|
| 🏛️ EU AI Act (2024/1689) | Articles 1–113, Annexes I–XIII, risk classification, provider & deployer obligations, GPAI, conformity assessment |
| 🔒 GDPR | AI training data, DPIAs, data subject rights, lawful basis |
Performance
| Format | RAM | Speed (CPU) | Speed (Apple Silicon) |
|---|---|---|---|
| MLX BF16 | ~2 GB | — | ~80 tok/s (M4 Max) |
| ONNX INT8 | ~1.2 GB | ~20 tok/s | ~60 tok/s (WebGPU) |
| GGUF Q4_K_M | ~1 GB | ~25 tok/s | ~90 tok/s |
Intended use
| Use case | |
|---|---|
| EU AI Act compliance Q&A | ✅ |
| GDPR guidance for AI systems | ✅ |
| AI risk classification assistance | ✅ |
| GRC documentation support | ✅ |
| Formal legal advice | ❌ |
| Autonomous compliance decisions | ❌ |
Target users: CISOs, DPOs, GRC consultants, legal counsel, compliance officers.
EU AI Act self-assessment
This model itself falls under limited risk (Article 50 — conversational AI system):
- Disclosure obligation: users must be informed they are interacting with AI
- No autonomous decisions — advisory only
- No personal data processed — on-device inference only
- Below GPAI threshold (< 10²³ training FLOPs)
Citation
@misc{allani2026euaiactedge,
author = {Allani, Sabri},
title = {EUAIAct-Qwen2.5-0.5B-Edge: A Sovereign Edge SLM for EU AI Act and GDPR Compliance},
year = {2026},
url = {https://huggingface.co/sallani/EUAIAct-Qwen2.5-0.5B-Edge},
note = {LoRA fine-tune of Qwen2.5-0.5B-Instruct on EU AI Act and GDPR corpus. Apache 2.0.}
}
EU AI Act · GDPR · Edge · Mobile · Offline · Open Weights
Apache 2.0 — Free to use, modify, and distribute