Locai Labs launches on-premises AI appliance for UK
Wed, 12th Aug 2026 (Today)
Locai Labs has launched Locai One, an on-premises artificial intelligence appliance built on NVIDIA infrastructure. The system is already being trialled by the UK Government's Incubator for AI.
The product is designed to let organisations run advanced AI models inside their own networks rather than through external cloud providers. It can be installed in an office or server room and connected to existing IT systems to support current AI applications and workloads.
The launch reflects growing interest among businesses and public bodies in keeping AI tools closer to their own data and infrastructure. That demand has been driven by concerns over security, intellectual property and the rising cost of cloud-based AI services that charge by usage.
Locai One combines NVIDIA RTX PRO 6000 hardware with Locai OS, an inference engine, governance tools and a new family of Juno models. The set-up is intended to provide a private large language model environment within an organisation's own network.
Locai Labs is led by founders James Drayson and George Drayson. James Drayson said the company's approach is aimed at making advanced AI small enough to run where data is held.
"To own and control AI deployment and accelerate AI adoption in regulated industries, we need to make advanced AI models small and efficient enough to run where the data is," said James Drayson, Chief Executive Officer at Locai Labs.
George Drayson pointed to the company's model compression work.
"Compress the model enough and the data centre becomes the AI box sitting in the corner of your office. Locai Labs' compression algorithms make it possible to move your AI in-house so you can own and control it," said George Drayson, Chief AI Officer at Locai Labs.
Model design
The first model being shipped on the system is Juno-N-Coder-25B, which Locai Labs said is based on NVIDIA Nemotron 3.5 Lightning, an open mixture-of-experts model built for reasoning, coding, planning and tool use. The company said it reduced the model's total parameter count from 31.56 billion to 25.57 billion using what it calls its SPACE algorithm.
According to Locai Labs, the process pruned experts by 25% and reduced the overall parameter count by 19% while retaining base-model performance on coding and agentic benchmarks including LiveCodeBench v6, HumanEval+, MBPP+ and SciCode. It said the trade-off was weaker performance outside its intended use, including general knowledge and factual recall.
The compression work is part of a broader effort to make large models usable on local hardware. Mixture-of-experts architectures activate only part of a model for each token, but deciding which experts can be removed without weakening a target domain remains a technical challenge.
Locai Labs said its SPACE method scores experts by their contribution to a chosen domain, in this case software development, and removes those judged least important. It then quantised the compressed model to NVFP4 for on-device inference on Blackwell GPUs.
Government trial
The UK Government's Incubator for AI is testing the appliance to assess whether sovereign, on-premises AI can support secure and data-sensitive workloads within government-controlled infrastructure. The trial gives Locai Labs an early public-sector reference point as officials examine where sensitive AI tasks should run.
"The Incubator for AI is pleased to be trialling Locai One and to support the kind of homegrown AI innovation that companies like Locai Labs represent. It is a great example of Britain's AI sector at work," said Max Hollingdale, Head of Applied AI Engineering at the Incubator for AI, UK Government.
Locai Labs said its UK sovereign AI assistant, GB1, already runs on a Jupiter-N-120B model built on Nemotron 3 Super and hosted on a Locai One NVIDIA GPU machine in the company's London office. It said that arrangement keeps the service on local infrastructure rather than in a remote cloud environment.
Cost pressure
Locai Labs is positioning the appliance as an alternative to cloud-based AI services whose costs rise as prompts, users and automated agents increase. It said an on-premises set-up offers a fixed hardware and software cost and removes token-based charges.
That argument comes as more organisations move from experimenting with AI chatbots to embedding models into internal workflows, coding tasks and document handling. At that stage, questions around governance, deployment and recurring operating costs become more pressing than they are at the pilot phase.
Locai Labs said it is developing further specialist models for scientific research, finance, law and healthcare. Its current focus remains on model compression, local inference and continual learning, with Locai One positioned as the main system through which those models are delivered.