Projects

What the lab
is building.

Each project tackles one piece of running frontier AI privately and efficiently. Some we release openly; some we take to market. When a project is done, or the world catches up, we say so.

Current project

AZ-Focus

Faster open frontier models: leaner reasoning and custom speculative decoders.

Two strands. Leaner reasoning: models spend much of their thinking on hesitation, so we optimise the weights against that habit and they answer with about half the thinking on reasoning and knowledge tasks, measured against the vendor's own effort settings. Custom speculative decoders: draft models tuned to different kinds of work, and designed to get faster as they learn a deployment's own traffic, with permission. Models released as open weights; drafters coming soon.

Results and releases →Hugging Face →
Research

Agent harnesses

Fitting open-source harnesses to open models and to real work.

An agent is a model plus a harness: tools, context, skills and guardrails. We fit open-source harnesses such as Hermes and pi to our models, and develop the method for fitting them to each business.

Read more →
Previous work

AiBrow

Local, on-device AI in the browser.

As browsers were looking to build AI in, AiBrow let web applications use local models privately on the user's own machine. Chrome and Chromium now ship built-in AI for the same job, so the project is now archived.

Previous work →GitHub (archived) →
In development

What's next

New open releases, and work we'll take to market.

More projects are in development. If you'd like to hear about them, or work with us on one, get in touch.

hello@axonzeta.ai →