01
It runs on your machines
Every system is self-hosted, and each can use a local model, through Ollama or a Cordon node, instead of an outside provider.
The mission
Why Dar es Salaam.
The tools on offer assumed English, a fast connection and someone else's servers. None of those can be assumed here, and much of the world shares at least one of the gaps.
So the models are trained on Kiswahili, on hardware a lab here can afford, and published with their evaluations. The software runs on an organisation's own machines, works with a local model, and keeps a record anyone can check.
- The working language is Kiswahili.
- The connection is not guaranteed.
- Compute is scarce and paid for.
- So the intelligence has to be owned.
Conditions
Built for where it runs. Each answered in code.
- 01The working language is Kiswahili.KW5 is small on purpose, for Kiswahili work where large multilingual models are too big to serve. It was trained on Kiswahili from the tokenizer up: 2.62B tokens across eight registers, from news and encyclopedia text to parliamentary records.
- 02The connection is not guaranteed.Cordon's Island and Dark modes run with no outbound connection at all, and take models on physical media. Every system can answer from a local model instead of an outside provider.
- 03Compute is scarce and paid for.KW5-149M was pretrained in about 27 TPU-hours on a Kaggle TPU v5e-8; the first generation trained on a single T4. Quantised to int8 it is 251 MB and runs on a laptop CPU. Every system can run on a local model.
Practice
Rules the code holds to.
- 02
The record can be checked
Cordon keeps a signed, hash-chained log that verifies offline, without trusting whoever runs the server. - 03
People approve what matters
Every Wallgarden tool starts at request-a-human, Matta's mappings wait for a steward, and Bubbly can open a pull request but cannot merge it. - 04
Limits are published
Model cards state what KW5 gets wrong. Product documentation says what has not been tested yet, such as Cordon's attestation on confidential hardware.
Leadership
The people leading the work.
Regnant's leadership, in Dar es Salaam, across the models, the machines and the deployments.
- Daudi IddFounder & Chief ArchitectSets the technical direction across the models and the systems.
- Aboubakar OmaryDirector of InfrastructureThe machines and networks the systems run on, from a single server to an isolated rack.
- Christopher LuogaLead Language EngineerKiswahili data pipelines, and the evaluations the models are measured by.
- McMillan ProterceChief Deployment OfficerTakes a system from the repository to a running installation on a client's own hardware.
Standards
Standards it answers to.
- PDPA 2022Data Protection, Tanzania
- e-GAe-Government Interoperability
- BoT Cyber RiskFinancial Data Isolation
- Ed25519 · SHA-256 chainTamper-Evident Logging
- OWL · SHACL · SKOSSemantic Governance
- Apache 2.0Open Weights: KW5