Regnant
Request access

Intelligence a nation can own outright.

Regnant builds AI models and institutional systems for infrastructure you control. Our work spans language research, secure inference, business operations, engineering and education, with ownership of the weights, deployment boundary and audit record kept within the institution.

Request access Explore the systems

Institutions don't shop for AI. They arrive with constraints that rule out the public cloud.

  1. 01

    Data that cannot cross the border.

    Private and secure inference with client identity, output policy, signed responses and an independently verifiable audit trail. Choose a deployment boundary and validate its hardware protections on the target platform.

    Cordon
  2. 02

    Decisions that carry liability.

    Connected business signals become evidence-backed briefs. You set the goals and operating limits; Wallgarden tracks approved work and measures claims against real observations.

    Wallgarden
  3. 03

    A language the stack was never built for.

    The working language of government is often not English, and translation breaks institutional precision. Our research holds the languages institutions govern in at the weights.

    Research

No borrowed ground.

From private inference to the classroom. Each system answers one institutional constraint, and each is handed over whole.

All systems

  1. Cordon

    Secure Inference Engine

    Private & secure inference. Verifiable answers.

    Open Cordon

Papers, engineering notes and dated releases, published when there is something to show.

Research NoteWhat would it mean for an institution to own its AI system?Model weights, local hosting and audit logs each solve only part of the ownership problem; this note traces the controls Regnant's systems add around them.Regnant Research, Christopher Luoga, McMillan Proterce
Blog EngineeringWhy we report bits per bytePerplexity cannot compare a 32,000-token Kiswahili vocabulary with a 256,000-token multilingual one. Bits per byte can, so it is the number the KW5 evaluation leads with.Regnant Research
Dispatch

Built in Dar es Salaam by engineers from the Dar es Salaam Institute of Technology and the University of Dar es Salaam.

Private & secure inference, with a record.
Checked at every step.

Cordon supervises local inference with client identity, output policy and signed responses. The deployment defines the perimeter; the audit trail lets you check what happened.

Cordon

A stack you can shape

Your institution

Start with the work. Connect the systems that belong together, then set the boundary around them.

01 / 03

Business and operations

Every decision has a path back to the work.

01KnottWorkflow02MattaSemantics03SeePOperations04WallgardenDecisions
02 / 03

Education

Orbis Teaching, study and research
Kairos Weave
03 / 03

Engineering and industry

01BubblyCoding agent02IINPlant floor
The inference boundary / 04CordonPrivate & secure inference.
Verifiable records.Explore the boundary
ON YOUR MACHINE

Pull a model

cordon pull HuggingFaceTB/SmolLM2-360M-Instruct-GGUF

Run it locally

cordon run smollm2-360m-instruct-gguf

Check the host

cordon doctor
Explore Cordon
Choose the systems your institution needs and the inference boundary that fits the work.
  • Supervised llama.cpp runtime
  • Per-client filtering and budgets
  • Offline signature and audit verification

Language research

Language is
infrastructure.

A system that cannot work in an institution's working language constrains the institution. We train models from scratch that hold the language at the weights, and publish the weights and the evaluation with them.

109M · 149M
KW5, two sizes, trained from scratch
Apache 2.0
Base and instruct weights, published
KW5 models and evaluations All research

Bits per byte on held-out Kiswahili. Lower is better.

Bits per byte by model, lowest first
ModelBits per byte
KW5-149M0.926
KW5-149M-instruct0.933
XGLM-564M1.122
BLOOM-560M1.699
GPT-2 124M2.126
Pythia-410M2.285
Qwen2.5-0.5B2.393

120 documents, 320,701 bytes, seed 20260914, run . Goldfish swa 100MB and Gemma 3 270M did not load and stay in the record.

The evidence

Choose your perimeter.

The boundary is set by your mandate. The system is configured to live inside it.

Discuss a deployment

  1. Air-Gapped

    Fully offline. No internet required. Models updated via secure media.

    Governments, central banks, defence, critical infrastructure.

  2. Private Network

    Runs on your LAN. Data stays on-premise; updates sync over a controlled tunnel.

    Standard enterprise and institutional deployments.

  3. Hybrid

    Core inference on-premise. Heavy training offloaded to a private cloud you control.

    Large organisations with dedicated IT teams.