Regnant
Search/Request access
NewIntroducing KW5-230M, our Kiswahili assistant

Sovereign AI · Dar es Salaam

Sovereign AI, built to understand Africa.

Models that learn our languages, and the way things are said here, from the first token, starting with Kiswahili. Systems that run inside your own perimeter. The weights, the keys and the record stay with you.

Explore the systems Research

L1Perimeter

Hardware you hold. Laptop to air gap.

Five deployment modes, from one laptop to a dark, air-gapped rack. A configuration that cannot meet its mode's rules is refused at startup.

L2Weights

Weights you keep. Kiswahili first.

KW5-230M learned Kiswahili from the tokenizer up, on text written here: news, parliament, the encyclopedia, conversation. It works sums step by step and remembers what you told it. Apache 2.0.

L3Cordon

One door to the model. Every answer signed.

Each request is written to the audit chain before the model sees it. The answer comes back signed, and the chain can be checked offline without trusting the operator.

L4Systems

The work on top. Owned outright.

Knott, Matta, Wallgarden and Bubbly each do one kind of work, run on machines you control, and can send their model calls through Cordon.

Skip the show
A standing stone cut with vertical grooves, on open moorland under cloud

02Sovereignty

Nothing in this stack is rented. The weights, the keys, the hardware spec and the audit record stay with the institution that runs them, inside its own border.

Why sovereign

03Language

Kiswahili, from the first token.

Igongwa ni kata ya Wilaya ya Kwimba katika Mkoa wa Mwanza, Tanzania.

16 tokens68 charactersVocabulary 32,000

A tokenizer learned from Kiswahili text keeps Kiswahili words whole. KW5-230M, the third generation, is small by design, trained here, and published with its evaluations and the limits it knows about.

Arithmetic, worked step by step
91.4%
Recall across a conversation
98.8%
The KW5 models Hugging Face

Sentence from Kiswahili Wikipedia, split by the KW5 v3 tokenizer. Figures: KW5-230M-instruct, frozen Kiswahili suite, October.

04Systems

Our systems. One perimeter.

CordonRuns your language model on your own hardware, and signs every answer.
KnottAutomates business workflows on your own servers, and sends the exceptions to a person.
MattaPuts your business records into one checked data model, and answers questions with sources.
WallgardenWatches your business systems against your goals, and proposes actions a person approves.
BubblyAI-assisted engineering with the rigor left in: every change planned, tested and reported.

Tell us the problem. We'll bring the system.

Request access All systems

New model · 29 October 2026

Introducing KW5-230M: Regnant's Kiswahili assistant.

It works a sum step by step, remembers what you told it ten turns ago, and says so when you never told it. 230M parameters, trained from the tokenizer up on 7.90B tokens of Kiswahili, open under Apache 2.0, and small enough to run on machines you own.

Meet KW5-230M Weights The report
Arithmetic
91.4%
bare sums, n = 128; the second generation scored 2.3%
Recall across turns
98.8%
a fact told 3 to 9 turns earlier, n = 80
Belebele
35.6%
the best of the small models compared, not separated from the next
Release bars
7 / 14
set before training; the missed ones are published
Mji mkuu wa Kenya ni upi?Mji mkuu wa Kenya ni Nairobi.Na wa Uganda je?Mji mkuu wa Uganda ni Kampala.
A follow-up from the evaluation run, unedited

Sovereign, in practice

Sovereign is a property of the code, not a promise.

Three things every system on this site holds to, each one checked in the code that runs it.

Research

Language is infrastructure. Kiswahili first.

The research behind the systems: models that understand the languages they are used in, measured in the open, served behind a door that signs what it says.
  1. 01ThreadKiswahili language modelsKW5-230M: tokenizer and weights trained on Kiswahili text, small enough to serve on modest hardware, released under Apache 2.0.The models
  2. 02ThreadEvaluation in the openEvery comparison ships with its manifest, its confidence intervals, and the bars it missed, published as missed.The evidence
  3. 03ThreadInference you can verifySigned answers and a hash-chained audit record, so an institution can prove what its model was asked and what it said.Cordon

From the lab

The newest work, as it ships.