Precision AI Get Maya
Researchers in 1930s work clothes run a laboratory in a hay barn. One tends tall cabinets of glowing vacuum tubes, one chalks a neural network beside wheat diagrams, and one types at a typewriter-keyed computing machine among specimen jars.

Research

A frontier lab for the oldest industry on earth

World-class researchers. One subject: food.

We build foundation models for agriculture, the edge computing that carries them into the field, and the machines they drive.

Model No. 1 · Reasoning

Beyond

Beyond is our reasoning model, trained for accuracy, speed, and token efficiency on agricultural work. It is the mind behind Maya.

Agriculture carries a high penalty for failure. A single wrong recommendation can put the wrong chemical on a crop, or the right one at the wrong rate, and take hundreds of thousands of dollars of value with it. Beyond is trained to abstain when it lacks certainty. That makes it safe to stand between a farmer and a spray decision.

A riveted brass computing machine in a barn with its service door open, revealing rows of glowing vacuum tubes and relays. An engineer on a step stool reads the paper tape while a man with a lantern looks inside.
020406080100PER CENT

53%

More accurate than frontier models on agriculture tasks

020406080100PER CENT

Abstention rate

Reading to be posted

020406080100PER CENT

Hallucination rate

Reading to be posted

Internal evaluation.

A vintage seed packet labelled SEED, Edge Model. A split seed sprouts a green shoot, with a glowing vacuum tube and brass gears at its core and copper wires for roots.

Model No. 2 · Edge

SEED

SEED is a multimodal foundation model for agriculture, pre-trained on more than 100 million images and optimized to run on the machine in the field.

SEED already understands crops, weeds, and soil, so a new use case needs as little as 10% of the data typically required. It turns raw imagery into compact plant-level vectors, taking terabytes of pixels down to kilobytes of meaning.

Models and datasets we have released

Core research · Edge computing

The field edge is where the work happens.

Farm machines generate more computer vision data than anyone can process offline. We run the models on the machine, in real time.

Our acceleration pipelines bring large models up to real-time speed on constrained NVIDIA hardware. The drone, sprayer, or tractor sees, decides, and acts in the same pass across the field.

Meet SOLV, our edge computer
Weathered hands hold a palm-sized finned aluminium module with a row of glowing tubes and a dial gauge over the open engine of a 1930s tractor.

Also on the bench

Eight more problems we are working on.

Secondary research that feeds the brain and the machines it drives.

  • 01

    Messy farm data

    Farm records arrive as scans, spreadsheets, PDFs, and photos of notebooks. Our pipeline reads all of it, with table extraction that outperforms almost all current OCR software.

  • 02

    Self-supervised learning

    A farm produces more imagery than anyone could label. Our models learn to ingest and classify it without labels.

  • 03

    Edge acceleration

    Pipelines that run powerful models at real-time speeds on constrained NVIDIA hardware, so vision data is processed on the machine as it is captured.

  • 04

    Real-time orthomosaics

    We stitch high-resolution orthomosaics while the vehicle is still moving, which removes the expensive offline processing step.

  • 05

    Vector compression

    We pull the meaningful content out of each image, vectorize it, and stream it back live with extreme positional fidelity.

  • 06

    Autopilot research

    Primary research into aerial and ground-based autopilot systems, with safe autonomy as the standard.

  • 07

    Variable-ratio fertilizer

    A hyper-precise sprayer that applies foliar nutrients in variable ratios, so every soil sample pays off in the field.

  • 08

    Herbicide resistance

    Models that detect herbicide-resistant strains, and species-level see-and-spray that puts higher-group herbicides only on resistant kochia and waterhemp.

Open and private research

Published where the field can use it.

We release datasets, models, and tools regularly on GitHub and Hugging Face, including AgriEmbedding, AgriStress-500, and MagniData. MCP layers and embeddings layers for agents are next.

Open releases Join the lab
Farmers in a 1930s Grange hall reading room study printed diagrams under amber lamps. In the middle of the table, two men work a typewriter-keyed computing machine whose round screen shows the same diagram.