Products
Research
News
Company
Careers
Search
Contact
Appearance
Arioron

We build the systems other systems depend on.

Arioron is a technology group operating across software engineering, applied research, network infrastructure and digital publishing. Five companies, one engineering culture, built to outlast the cycle.

Scroll
Trusted by teams at
  • Maximoft Ltd
  • Drape & Dash
  • Octagon Learning
  • Project Compassion
  • Tonger Khobor
  • EGMZ Community
01 — Figures

The group in figures.

Figures are stated with the context that makes them mean something. We would rather publish a small honest number than a large decorated one.

5
Operating companies

Each with its own leadership, market and standards.

2022
Founded

In Dhaka, Bangladesh. Engineer-led since.

6
Models published

Open weights on Hugging Face, with evaluation code.

0
Hyperscaler dependencies

The group runs on hardware Vortarz Network operates.

02 — The group

Five companies.

Each company runs its own practice and answers to its own market. What they share is an engineering standard and an infrastructure spine.

Structure

Holdings owns. The companies operate.

One ownership entity, four operating companies, and a shared spine of engineering standards and infrastructure.

Arioron Holdings Ownership · capital · horizon Arioron Developers
Software engineering practice
AD Arioron Labs
Applied research
AL Vortarz Network
Game development and server hosting
VN Elakar Khobor
Local news platform
EK Shared: engineering standard, infrastructure spine, publishing platform
03 — In focus

Arioron Labs

Applied research

Intelligence amplified.

Cutting-edge AI research lab creating the next generation of intelligent systems — including the VexAI assistant and advanced language models published openly for anyone to check.

Arioron Labs
Founded
2024
Published models
6
Open datasets
3
Largest model
2B parameters
Licence
Open weights
04 — Approach

Infrastructure, not features.

Most technology work is disposable. It ships, it drifts, and it is rewritten inside three years.

We are interested in the other kind — the systems that quietly hold up everything else. That means owning our infrastructure rather than renting it, writing software we are willing to maintain for a decade, and publishing research openly so it can be checked.

It is a slower way to build. It compounds.

05 — Arioron Labs

Published research.

ARL-2025-011

Vex-Amber-Fable-2.0: parameter efficiency in the sub-3B class

A 2-billion-parameter causal language model trained for software engineering and reasoning tasks, evaluated against models an order of magnitude larger. We report...

Read the paper
Results ARL-2025-011
SWE-bench (Verified) Agentic software engineering
65.37%
HumanEval Function-level code generation
60.98%
LiveCodeBench Contamination-resistant coding
44.19%
AIMLE reasoning Multi-step reasoning
51.39%
ARL-2025-009 OncoDetect Titan: robustness over benchmark fit in lung cancer detection A multi-model ensemble for lung cancer detection, trained explicitly on degraded and inconsistent scan data. We argue that robustness to real hospital conditions matters more than... Jul 2026 ARL-2025-004 Vex-Amber-Mini: reasoning at 0.6B parameters A 0.6-billion-parameter model optimised for code generation and general text tasks on constrained hardware, with notes on the distillation approach and where it fails. Apr 2026
Open source

Published, so it can be checked.

Hugging Face

Arioron Labs releases model weights, training datasets and the evaluation harness together. A benchmark result nobody can reproduce is marketing, not research.

Everything is on Hugging Face under a permissive licence, including the models that did not work as well as we hoped.

Weights

Published for every released model.

Datasets

The training data, not just a description of it.

Evaluation

The harness that produced the numbers.

Failures

Documented as carefully as the wins.

Infrastructure

We own the machines.

Vortarz Network runs bare-metal compute, storage and routing for the whole group. Nothing here sits on a hyperscaler by default.

That decision sets our cost floor, keeps client data off third-party infrastructure, and means an incident is ours to fix rather than ours to report.

2B
Largest model

Parameters in Vex-Amber-Fable-2.0.

2
Independent routing paths

So the loss of one is survivable.

3
Open datasets

Released alongside the models trained on them.

24/7
Named on-call

A person, not a support queue.

Working together

How an engagement runs.

The same shape whether the work is a platform build, an infrastructure engagement or a research collaboration.

01

Scope

A short paid discovery. We write down what the system must do, what it must never do, and what we are choosing not to build.

02

Architecture

The shape of the thing, in writing, before any code. You get the trade-offs we considered and why we landed where we did.

03

Build

Two-week increments against a running system. You see it working, or not working, the whole way through.

04

Operate

We stay responsible for it. Most of our work is with clients we first shipped for years ago.

06 — Company

A group, not a holding pattern.

Arioron began in Dhaka as a small engineering practice and grew into a group of five companies. We are engineer-led, remote-first, and deliberately independent — our infrastructure runs on hardware we operate ourselves.

That independence is not ideology. It is what lets us make long commitments to the people who build on top of us.

About Arioron

Safwat Shabib

Chief Executive Officer

Mashrur Hassan

Chief Marketing Officer

Md Shafayet Shayem

Chief Of Operations

Md Tahmid Islam

Senior VP, Graphics

Ming Wei

Senior VP, Healthcare AI

Mubashirul Islam Yasin

Chief Technology Officer

07 — Careers

Open roles.

We hire people who want ownership of real systems and are willing to still be maintaining them in three years. Small teams, direct responsibility, no theatre.

All roles

Build something that is still standing in ten years.