← All publications

2026-07-20

Pyxon SLLM v0 — Arabic-first small language model for edge and on-device AI

Hamza Salem, Ahmad Algraeeb, Almuhanad Al-Nihmy & Manuel Mazzara

Ollama · edge AI · Arabic · QLoRA · on-device · SLLM · Qwen

A ~1.5B Arabic instruction model fine-tuned for chat and edge deployment. Run offline with Ollama — pair with on-device RAG when accuracy must match larger cloud systems.

Pyxon SLLM v0 · Arabic-first small language model for edge and on-device AI
Built by PYXON.AI · Research · Labs · Production AI · Beta · v.2026

Why pay for a 7B cloud model when a 1.5B model on your device can carry the workload — especially with retrieval? Try PYXON.AI →


What this is

Pyxon SLLM v0 is a ~1.5B Arabic instruction model fine-tuned for chat and edge deployment. It is part of PYXON AI Labs’ on-device research line: lightweight models that run offline, keep data on-device, and pair with retrieval (RAG) when accuracy must match larger cloud systems.

Name pyxon-sllm-v0
Base Qwen2.5-1.5B-Instruct
Focus Modern Standard Arabic instruction-following
Method QLoRA supervised fine-tuning on public Arabic datasets
Runtime artifact pyxon-sllm-v0.gguf (q8_0)
Designed for Phone / edge / Ollama / offline assistants

From the Labs

PYXON.AI is PYXON’s research division — scientists and engineers building AI that performs on Arabic-language and on-device challenges, not slides.

This release sits next to Pyxon On-Device RAG: research showing that small models plus device-specific retrieval can match or beat much larger baselines while staying compatible with mobile compute, memory, and privacy constraints.

Enhancing weak LLM performance on edge devices through retrieval-augmented generation — a benchmark study (Beta · 2026-03-28)

Hamza Salem, Ahmad Algraeeb, Almuhanad Al-Nihmy & Manuel Mazzara

A 1,500-question benchmark from three phone manuals shows RAG lifting small on-device models to 100% accuracy versus ~21–37% without retrieval, with analysis of size, latency, and device-specific knowledge bases.

Hamza Salem — Head of Pyxon AI Department, PYXON.AI (hamzas@pyxon.ai)
Ahmad Algraeeb — PYXON AI Department (ahamdg@pyxon.ai)
Almuhanad Al-Nihmy — Kütahya Dumlupınar University
Manuel Mazzara — Innopolis University

More products and labs updates: www.pyxon.ai


Quick start (Ollama)

Pull and run from Ollama Hub:

ollama run enghamzasalem77/pyxon-sllm-v0

Or build locally — place pyxon-sllm-v0.gguf next to Modelfile, then:

cd pyxon-sllm

# rename if you still have the old export name
# mv arabic-1.5b-q8_0.gguf pyxon-sllm-v0.gguf
# mv outputs/gguf/pyxon-sllm-v0.gguf ./pyxon-sllm-v0.gguf

ollama create pyxon-sllm-v0 -f Modelfile
ollama run pyxon-sllm-v0 "اشرح لي الذكاء الاصطناعي باختصار"

Model card: ollama.com/enghamzasalem77/pyxon-sllm-v0

Name Size Context Input
pyxon-sllm-v0:latest 1.6 GB 32K Text

Training data (public)

Dataset Role
arbml/CIDAR Culturally reviewed Arabic instructions
FreedomIntelligence/alpaca-gpt4-arabic Broad Arabic SFT
CohereForAI/aya_dataset (Arabic) Multilingual quality filter
FreedomIntelligence/evol-instruct-arabic Harder instructions (optional)

Reproduce the fine-tune

cd pyxon-sllm
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

python prepare_data.py --include_evol
python train.py --data data/train.jsonl --eval_data data/val.jsonl
python merge_and_export.py --quant q8_0
# if merge already done:
# python merge_and_export.py --skip_merge --quant q8_0

Requires a GPU with ~16 GB VRAM for QLoRA.


Intended use

  • Arabic chat / Q&A on edge devices
  • Offline or privacy-sensitive assistants
  • Backbone for on-device RAG (retrieve from local docs, then generate)

Not a substitute for specialist medical, legal, or safety-critical systems without human review.


GPU fit checker

Not sure which LLM sizes your hardware can run? Use GPU Lab — paste your terminal GPU output, match to our catalog, and see inference vs fine-tuning capacity for models from 1B to 405B.


Research that actually works — Arabic-first, on-device, shipped.

www.pyxon.ai
→ Join Labs / careers: info@pyxon.com · hr@pyxon.com