We design compact AI that runs on the hardware you already have, from robots to sensors to phones. No datacenter required.
Most frontier AI assumes a datacenter behind it. We think there's another way.
We're an independent lab working on AI that runs at the edge. Most of our research starts from one question: how much can a model do when it has to fit on the device in front of you?
We design for constrained hardware from the start, choosing architecture, quantisation and distillation around where a model will run. That makes advanced AI practical for the teams, devices, and regions that large-scale infrastructure leaves out.
We're early. We share our work as results arrive.
Build AI that performs at the edge, with capability driven by the quality of its design rather than the size of its compute budget.
A future where the most useful models are the ones you can run yourself.
Corrective learning in place of scale for efficient manipulation.
A task-conditioned vision-action policy trained from scratch. No language model. No pretrained backbone. Robustness comes from the training signal, rather than the parameter count.
¹ Nearest peer: 0.821 mean success. Results as reported in the paper.
Control and perception policies that run on-device, for robots and machines that can't phone home for every decision. PRAXIS and SEAL are our first result in this direction.
First results · Accepted at NeurIPSOur 2.01M-parameter vision encoder works without a pretrained backbone. We're developing it into a family of compact models for retrieval, perception, and understanding.
In progressA full-stack coding and research agent for hardware you already own.
In its flagship mode it runs entirely on one 8 GB consumer GPU: no API keys, no metering, and no request leaves the machine. The same harness can also drive Ollama, LM Studio, or a cloud provider.
Calibrated, typed decisions from open LLMs. No text generation.
You send a state and a set of typed questions. It returns a calibrated probability for each option, plus the probability that none of them fits. It runs locally, down to a Jetson.
Write together. Compile anywhere.
A self-hosted collaborative LaTeX editor in a single binary: live editing that keeps working offline, sandboxed builds, git-backed history, and a project-scoped MCP server for the AI client you already use.
Very long contexts for a small model, on modest hardware.
Mercurius rebuilds the memory-hungry attention of Qwen3.5-4B into a compact latent form, then distills it from a larger teacher. It reads a 1M-token context in about 5 GB instead of about 37 GB, and runs in llama.cpp.
Talks, guides, and reference material from our own work. Free to take and adapt.
From the fundamentals to an assistant that works over your team's documents, code, and data. It covers retrieval, harnesses, local inference, fine-tuning, and verifiers, with a phased roadmap.
Print or save as PDF from inside the guide.
Send us a few lines about what you're building and the hardware it has to run on. We'll reach out.
contact@minervalabs.mx