Intelligence is inevitable. Trust is not.
nowara is a UK AI company. We map how AI is built, deployed and used, and test whether it can be trusted. AICAP, our first product, does this for organisations. We also train, fine-tune and quantise models for specific tasks, and research how AI systems are attacked and how they are used in attacks, so we can defend against both.
products
-
AICAP
Continuous AI assurance, compliance and security across the AI lifecycle. It works out what to test, generates the tests and keeps the evidence. It has four parts, and includes a model arena for comparing models before you buy or deploy them.
- Survey discover
- Maps every model, agent and tool in use, from what endpoints and coding tools report.
- Origin build
- Records which AI model wrote which code, and re-checks that code when a model is found to be flawed.
- Core deploy
- Assures deployed models and agents as their use changes, and generates the tests that matter before go-live.
- Trace operate
- Watches what agents do in production: the tools they use, the steps they take and the actions they carry out.
-
Model Registry
A public, continuously updated record of AI models: releases, variants, quantisations and formats, with a source for every entry. It does not score or rank models.
-
Next
Our own task-specific foundation models, fine-tunes and quantisations, and AI security tools from our red-team research. We will list them here when they are ready to use.
model registry
Look up any model, its quantisations and where they run.
A single LLM release can lead to dozens of downloadable files: different formats, quantisations and fine-tunes. The registry links each file to the model it came from.
| model | publisher | variant | format | quant | params | size | released |
|---|---|---|---|---|---|---|---|
| Gemma 3 12B | it | GGUF | Q4_K_M | 12.2B | 7.3 GB | 2025-03-12 | |
| Mistral Small 3 | Mistral AI | Instruct 2501 | GGUF | Q4_K_M | 23.6B | 14.3 GB | 2025-01-30 |
| DeepSeek-R1-Distill | DeepSeek | Qwen 7B | safetensors | bf16 | 7.6B | 15.2 GB | 2025-01-20 |
| Phi-4 | Microsoft | base | GGUF | Q6_K | 14.7B | 12.0 GB | 2024-12-12 |
| Qwen2.5 7B | Alibaba | Instruct | AWQ | int4 | 7.6B | 5.6 GB | 2024-09-19 |
| Llama 3.1 8B | Meta | Instruct | GGUF | Q4_K_M | 8.0B | 4.9 GB | 2024-07-23 |
- Sourced
- Each entry links to where it came from: the model card, licence, weights metadata or release notes.
- Up to date
- New releases, re-uploads and community quants are added when they are published.
- No rankings
- The registry lists models without scoring them. To compare and evaluate models, use AICAP.
about
About nowara
Nowara AI was founded by Nick, Usman and Ciaran. We started it because we could see both sides of AI: how much good it could do for people, and how easily it can fail or be misused. We want to help people get those benefits without taking on risks they can't see.
Most organisations now use AI models they did not train, often in versions the original developer did not publish, and those models change more often than teams can review them. We work on three things: AI models, AI assurance and AI security. Part of that work is attacking AI systems, because it is how we find out where they fail.
The name comes from the Arabic root n-w-r (نور), the root of noor, meaning light. Nowara means a single flower, a bright white blossom. We liked both meanings: light, because our work is about making AI easier to see into and understand, and a flower, because we want what we build to be good for people.
Our logo is a sunflower head drawn in code: 200 dots, each turned about 137.5 degrees from the one before, which is the arrangement sunflower seeds follow. Nowara AI Limited is a UK company.