AI product and research company Β· UK
Your AI, on hardware you own.
vocabotics is a UK AI product and research company. We build NovaTerra, a personal AI that runs on hardware you own, and Quanta, a verified GPU inference compiler for it.
Lab result
Re-run in progressQuanta decoded 1.27Γ faster than llama.cpp, in one narrow case.
Qwen2.5-0.5B Q6_K, batch-1 greedy decode, one RTX 3090, same session: 844 against 661 tokens a second, with every tokenβs logits bit-identical to its CPU reference.
Where it does not hold: at 7B our engine was 0.88Γ llama.cpp, and prefill is slower. Measured 5 August 2026; a re-run with full logs is in progress.
vocabotics builds personal AI that runs on hardware you own, and the GPU compiler underneath it. Every result comes with the proof to check it.
We say what was measured and where to check it. We say what isnβt proven.
What we make.
Two products and the platform we run on, each with its real status.
- Our main productNovaTerraPrivate beta, invite-onlyA personal AI operating system. Each person runs their own node: a Twin that keeps their memory and files, a Muse they can talk to, and a Studio where agent crews build documents, code and sites.How NovaTerra works β
- From our labQuantaResearch prototypeA compiler that turns a model into PTX for NVIDIA GPUs, with every kernel checked bit for bit against a reference. It is how NovaTerra will run on a GPU its owner controls.The result and its limits β
- The platform we run the company onvocabotics ConsoleIn daily use here, not on saleOur own operating system for the company: AI staff do the routine work, a person says OK before anything leaves, and everything lands on one record we can check.Read more β
The Lab.
The research NovaTerra needs to run on a GPU its owner controls. Dated results, failures included.
- Quanta on NVIDIA GPUsVerified PTX, one CUDA graph per token, measured against llama.cpp on an RTX 3090.Read β
- Wave-ternary modelsA 278M-parameter pilot trained on one RTX 3090. No win over a dense model yet; the matched control is next.Read β
- The Lab NotebookEvery dated lab report since 2024, failures included.Read β
Who runs it.
Jon Ross, founder and director, and the full-time developer. Fifteen years engineering safety-critical rail and metro systems, then four years of AI research, and now vocabotics. AI agents do build work under his review. Meet the founder
Write to jon@vocabotics.com. Jon replies personally.