Independent research. Intelligence at the edge.

Intelligence,
closer to the work.

We build efficient AI for the machines, devices, and people doing the work. Designed for real hardware. Capable by design.

Explore our research Meet Minerva
From the lab 01 / PRAXIS
7.56M

Parameters. A different approach to capable robots.
01 / The lab

Capability shouldn't depend on a compute budget.

Most frontier AI assumes a datacenter behind it. We think there's another way.

We're a small, independent team building efficient, economical models for edge deployment, from embodied and physical AI to compact models for retrieval, perception, and understanding.

We design for constrained hardware from the start. 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.

Our mission

Build AI that performs at the edge, with capability driven by the quality of its design rather than the size of its compute budget.

Our vision

Capable AI embedded in the machines, devices, and tools around us, available to anyone who wants to build with it.

02 / Research

More from less.

Featured researchDraft · Under review

PRAXIS over Theoria

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.

With SEAL, we roll the policy out and supervise every state it visits with an analytic recovery action derived from the demonstration itself. No interactive expert required.

7.56Mtotal parameters
Mean success · LIBERO
0.831
Smaller than nearest peer¹
9×
Vision encoder parameters
2.01M

¹ Nearest peer: 0.821 mean success. Results reported in the draft; under review.

Next: a standalone family of compact vision encoders for retrieval, detection, and understanding at the edge.
01

Embodied intelligence

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 · Paper under review
02

Compact representations

Our 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 progress
03

Efficiency & deployment

Training and inference methods that treat the hardware budget as a design input. Quantisation, distillation, and architectures chosen for the hardware they'll run on.

Ongoing research
03 / Product
Open source · In development

Promethean

View on GitHub

A capable coding agent for hardware you already own.

Promethean began as a full-stack local agent. As its layers multiplied, the system grew away from the idea it was built to prove: capable AI can be small, understandable, and efficient.

We rebuilt it around a smaller core. The repository is currently called MiniHarness; the product remains Promethean while we develop it in the open.

Implementation
~6.6K lines
Tests
~5.6K lines
Tools
8
Core dependencies
0
The lesson became the product: every tool and every token must earn its place.
04 / Writing

From the workbench.

Papers & preprints

2026 / Research paper

PRAXIS over Theoria: Corrective Learning in Place of Scale for Efficient Manipulation

Under review

The preprint will be linked here when it becomes available.

Notes in progress

Why we design models small from the start

Draft

A compute budget is a design constraint, not a ceiling

Draft
05 / Resources

Built to be shared.

Talks, guides, and reference material from our own work. Free to take and adapt.

Technical guide45 slides · English

LLMs on your own corpus

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.

For
Research & engineering teams
Duration
One hour + reference appendix
Prerequisites
None
Open the guide

Print or save as PDF from inside the guide.

06 / Contact

Good work starts
with a conversation.

Building on constrained hardware? Exploring a research collaboration, a pilot deployment, or an investment in efficient AI? We'd like to hear from you.

contact@minervalabs.mx