TL;DR: TechnologyKSA has released AbdoCode, a Saudi-built AI coding agent that runs on your own Windows machine and works with 17 built-in model providers or with local open-source models through Ollama, so code never has to leave the device. We measured it on a 2-billion-parameter model and on larger models, and we publish the results as they are — including the limits of small models.
What is AbdoCode? AbdoCode is an AI coding agent developed by TechnologyKSA that runs on Windows 10/11 with a small, deterministic Rust kernel. It writes, edits and runs code inside your project, shows every change before it is applied with rollback, and works either with a local model on your machine or with your own API key at a cloud provider.
Why AbdoCode is a Saudi first in AI agents
Most well-known coding agents are designed around their vendor's own models. AbdoCode takes the opposite approach: the model is your choice. The current public release (4.0.23, published on GitHub on September 14, 2026) ships a catalogue of 17 built-in providers and automatically tightens its safety rails according to model size — the smaller the model, the stricter the guards on writing and executing.
Full privacy: when does your data stay on your machine?
Direct answer: in local mode, AbdoCode talks to an Ollama server at 127.0.0.1 on your own computer, so neither code nor instructions leave the device. When you choose a cloud provider, requests go only to that provider, with your key, and the key is stored encrypted in the Windows Credential Manager (DPAPI). No TechnologyKSA cloud account is required.
From 2B parameters to large models: what we actually measured
We ran a real coding task on an isolated copy of the installed engine: create a sum function, a test file, then run node test.js. We verified every result by running the test ourselves, outside the agent.
| Model | Where | What happened | Verdict |
|---|---|---|---|
| Qwen 3.5 — 2B | Local (Ollama) | 22 tool calls in 28 seconds: wrote both files, the first run failed with a TypeError, it read the error, fixed it and re-ran with exit code 0. | Drives the agent loop, but its final test copied the function instead of importing it and printed success unconditionally — needs supervision. |
| Qwen — 27B | Local (Ollama, 75% on CPU) | Only 6 tool calls: an exported sum.js, a real node:assert test that imports it, PASS with exit code 0, and an accurate summary. | Completed the task exactly as specified — slower on limited GPU memory. |
| Nemotron 3 Super — 120B | Cloud (NVIDIA NIM) | A real Blender project: villa.blend, four 1280×720 renders, a 756 KB PDF and an honest README of what was left undone. | Handles multi-step work; the scene quality itself was weak, and the report said so. |
AbdoCode and HUMAIN's humain-m3
humain-m3 is an Arabic model announced by Saudi Arabia's HUMAIN on September 3, 2026 at LEAP in Riyadh: a mixture-of-experts model with about 428B parameters (23B active), built on MiniMax-M3 and trained on over a trillion Arabic tokens, currently offered as a research preview on HUMAIN Node through an OpenAI-compatible API. AbdoCode accepts a custom cloud provider on any OpenAI-compatible https endpoint, so connecting it needs no code change. To be clear: we verified this in AbdoCode's validation code; we have not yet run a live session on humain-m3.
How does AbdoCode compare with Claude Code and Codex?
Claude Code (Anthropic) and Codex (OpenAI) are among the strongest coding agents in the world. We have not run a benchmark against them and we do not claim superiority we have not measured. What we can show today is a different philosophy: free choice of model, fully local operation, Arabic-first interface, adaptive guards, vision tools and a three-lens self-review.
What it means for Saudi companies
Teams that cannot send source code or customer data to external servers can now run a coding agent on their own machines, with the model they choose. Personal use is free; commercial use requires a written licence from TechnologyKSA — contact us for enterprise programmes and partnerships.
Frequently asked questions
Is AbdoCode free?
Personal use is free. Commercial use requires a written licence from TechnologyKSA.
Does AbdoCode work offline?
Yes, with a local model through Ollama; traffic stays on the machine at 127.0.0.1.
What is the smallest model it works with?
We measured a 2-billion-parameter Qwen model: it wrote, ran and fixed code, but its results need human review. Larger models are more reliable.
Which operating systems are supported?
Windows 10 and 11, 64-bit.
Sources
- AbdoCode public repository and releases — github.com/code-ksa/abdocode (v4.0.23, September 14, 2026).
- humain-m3 announcement — Unite.AI and Bloomberg (September 3, 2026).
- AbdoCode measurements in this article: runs by TechnologyKSA on September 14, 2026, on installed build 4.0.24 with an isolated engine copy and independent output verification.
