Independent AI lab · Open weights · Manchester, UK
Frontier AI, where the data is.
We help organisations run powerful AI privately, on their own hardware, without sending their data to anyone else.
AxonZeta is an independent AI lab working on open weights. We build the pieces that let frontier AI run privately and efficiently: leaner open models, agent harnesses fitted to real work, and the tools to measure them. Some projects we release openly; some we take to market.
Start from an open model, then add what makes it yours: your documents, your tasks, your tools and your usage. It runs on your hardware and keeps improving.
height = capability
Your private modelv1.0
Drafteryour usage
Harnessyour tools & rules
Fine-tuningyour tasks
Continued pre-trainingyour documents & code
Open-weight baseDeepSeek-V4.1
ours, built on site open-weight base, as shipped
Built on open models
DeepSeekGLMQwenKimiMiniMaxNemotronInkling
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.
Open-weight models are now close to the frontier. Our research is about what it takes to make them genuinely useful outside a hyperscaler: cheaper to run, faster, better at a specific business, and able to keep learning.
GenericKnows your business
01Active
Domain adaptation
Continued pre-training and fine-tuning of frontier open models on a domain's documents, code and tasks, and how to do it without losing the general ability that makes them useful.
JanAprJulOctNow
02Exploring
Continual learning
Models that keep up as new data arrives, without forgetting what they knew.
YourQ3churnreportisshowsa2.1%drop
drafted & acceptedcorrected
03Active
Faster inference
Models write several words at once by letting a small draft model guess ahead. As part of AZ-Focus we build custom speculative decoders (draft models) tuned to different kinds of work and retrained on a deployment's own traffic, with permission. We serve frontier models on NVIDIA clusters, from Sparks to GB300.
Before
888
AZ-Focus
339
thinking tokens per reasoning task (BBH)
04Shipped
Leaner reasoning
Reasoning models spend a lot of time second-guessing themselves. We optimise the model's weights against that habit: it thinks less, answers sooner, and keeps its accuracy. Released for DeepSeek-V4.1-Flash and GLM-5.3-Flash.
When a reasoning model "thinks", much of it is hesitation: "Wait…", "Hmm…", "Let me double-check…", re-working answers it already had. Every one of those words costs time and compute. Our AZ-Focus models keep the reasoning and drop the hesitation.
QuestionA train leaves at 3:40 pm and arrives at 6:15 pm. How long is the journey?
Original model0 tokens · 0.0s
AZ-Focus0 tokens · 0.0s
reasoning AZ-Focus reasoning scrap: "wait", "hmm", "let me double-check…"
Illustration: each block is about 10 thinking tokens. The lengths match our measured averages on reasoning benchmarks (about half the thinking tokens) with accuracy within about three points. Fewer thinking tokens means less to generate, so answers generally arrive sooner.
2.2×
less thinking on general knowledge (MMLU-Pro)
2.6×
less thinking on logic & reasoning (BIG-Bench Hard)
2.0×
less than DeepSeek's own lowest effort setting
≤3
points below the original at most on knowledge & reasoning; lower on competition maths
DeepSeek-V4.1-Flash-AZ-Focus vs the original, same serving and settings.Full results →
Our thesis
We build the pieces private AI runs on.
Most businesses can't, or shouldn't, send their most valuable data to someone else's model. Open weights change that, if they can be made efficient enough to run on-site or in a private cloud, and specific enough to be worth it. That's the gap we work on: the model layer for private AI.
We're based at Sister, Manchester's innovation district, and work on real hardware: NVIDIA clusters from Sparks to GB300. They run frontier models with hundreds of billions of parameters, and every result we publish is measured on them.
Before this, we built AiBrow, which brought local, on-device AI to the browser as browsers were looking to build it in. Chrome and Chromium now ship that natively, so AiBrow is now archived; the idea of running AI where the data is carries on here.
Open weights: inspectable, auditable, never changed under you
On-site or private cloud, inside the security boundary
Measured in the open, with results we publish
Work with us
Building on open models? Let's compare notes.
Our models are free to use. If you're running them, testing them or building on them, we'd like to hear what you measure.
A small number of organisations working with us directly on hard problems
Early access to our models and methods
With your permission, results we can publish
Releases · AZ-Focus
Open models
If you only have a minute, which gets you the better answer: the original model at a lower effort setting, or a focused model thinking harder? Same hardware, same time budget. AZ-Focus models are built to explore that trade, and we publish what we find, including where it doesn't hold.
Open weights, measured and documented, free to use.