Building the first Large Geospatial Model to answer the most difficult questions about our planet.
- Contact
- contact@columbus.earth
- Social
Building a brain for earth.
At Columbus, we collect the world’s data, and build a brain that comprehends it all.
We’re building frontier geospatial intelligence.
An interactive Thesis
Large Geospatial Model vs Large Language Model.
If an LLM is for the digital world, our LGM is for the physical world.
We’re personifying earth with physical AI. Instead of words, we process data about our surroundings and the anthropology in them.
A new foundational model is needed.

An LGM vs other foundation models
LLM
Large-Language-model
VLM
Vision-Language-model
LGM
Large-Geospatial-model
Text
e.g. “The grass is” → “green”
Text & Image
e.g. dog photo → “a border collie”
Physical reality
e.g. public data + urban imagery + GIS → crime risk map
outputs
Predictive words
“What word comes next?”
Visual reasoning
“What’s in this image?”
Ground truths
“What’s in this physical space?”
building it







Columbus EarthA Large Geospatial Model is the next frontier in AI
The path forward
& Vision Models

LGM
Timeline of foundational AI models
- 2022 — LLM (Large Language Model)
- 2025 — Geo-tuned LLM & Vision Models
- 2027 — Generalist LGM (Large Geospatial Model)
- 2028 — UGM (Universal Geospatial Model)
- 2022LLM
- 2025Geo-tuned LLM & Vision Models
Now — July 2026- 2027Generalist LGMRead our Paper
- 2028UGMOur Game Plan
the drawbacks of LLMs & vision models.
How the LGM innovates.

Our Reasoning
We've come up with several innovations within data collection, fusion, and core reasoning. We combine several innovations in unique ways in our research.
We learned first-hand how LLMs are not built for geospatial needs. We set out to fix every technical issue that came with GPT architecture and it converged into a new foundational model, the LGM.
Our proprietary architecture is comprised of 3 parts. Within each are several innovations built during our practical research.
The most extensive data collection in the industry. Versatile methods ranging from drones, car data, human data, public data and more.
We've achieved the cheapest P/POI.P/POI — Price per Point of InterestThe cost to capture a single geospatial data point. Lower P/POI means richer, more affordable training data — a key economic metric for any geospatial foundation model.
Read our blogAccurate, automatic data filtering & labeling.Automatic labelingThe process of tagging raw, unstructured geospatial data with structured metadata so a model can reason about it. Our pipeline does this without human annotation.
We care about Ground Truths.Ground truthA verified, real-world observation confirmed at a specific X, Y, Z coordinate and time. Ground truths are the reference signal we use to validate every layer of the model. We make sure each data point is the truth at that specific X,Y,Z point at that given time.
Data scarcity is one of the hardest parts about the LGM endeavor. To solve this, we have built innovative methods to universally digest data. Meaning we are able to fuse data together that our model then trains on. Cheaper and more data → smarter model.
Our reasoning model considers temporal data, and sifts through vast amounts of aggregated geospatial data — including anthropologic data.
It continuously learns and creates new patterns through our relational architecture.
Our core reasoning is comprised of a new permutation of Reverse Diffusion and RAG architecture.
Read our PaperData Collection
The most extensive data collection in the industry. Versatile methods ranging from drones, car data, human data, public data and more.
Read our blogour foundational model research
Our Model: Magellan-1.0
The latest results from our development of the LGM.
Contextually enriched reasoning over the semantics of urban space
Generative Geospatial data
A generalist model, grounded on a living data catalogue
Granular reasoning at scale
Read our latest releases
Explore the innovative research and recent papers from our team.

Join us! Research opportunities for the rebellious
Columbus Earth is hiring — open research opportunities for those who want to build the Large Geospatial Model and help shape the future of geospatial AI.

The Large Geospatial Model: Our Timeline
Neque porro quisquam est, qui dolorem ipsum quia dolor sit amet, consectetur, adipisci velit, sed quia non numquam.

Turning Dormant Data Usable
Nemo enim ipsam voluptatem quia voluptas sit aspernatur aut odit aut fugit, sed quia consequuntur magni dolores eos.
Careers
If you're excited about creating paradigm shifts in physical world understanding.
Join the crew
Research freely at Columbus.