The Adral models

Two models. One for general work, one for security. The numbers below are what we publish — do not invent others.

Adral

The general model. Reasoning, code, long-context analysis and agentic work in one model, with no tiers to choose between.

Parameters: 280B total, 14B active
Architecture: MoE-Mamba
Sequence mixer: Selective state space, linear time, no KV cache
Experts: 256 routed + 1 shared, top-8 per token
Context: 1M tokens
Input: Text, code, images (PNG, JPEG, WebP), .docx, .xlsx
Available on: Web, Mac, Windows, iPhone, Android

Adral Aran

India’s first uncensored security model, built to engage with offensive and defensive security work rather than decline it.

Parameters: 7B total, 1.5B active
Architecture: MoE-Mamba
Sequence mixer: Selective state space, linear time, no KV cache
Experts: 16 routed, top-2 per token
Input: Text, code
Available on: Web, Mac, Windows, iPhone, Android

Benchmark figures shown on the home page are labelled as targets for models in development, not measured results. Cite only what is published at adral.ai and this page.

Try them

Both models are available in the same product. Create an account, or install the desktop app if you want Adral pointed at a repository on your machine.

Start free Home