Geospy coined the term Large Geospatial Model. Orca 1.5 is Oceanir's version of the idea: a model that works out where something is, which way the camera was pointing, and what is around it.
Most AI is good at telling you what a photo shows. A person, a car, a dog. Orca 1.5 is aimed at a different question: where was this taken, and how sure are we?
Geospy gave this kind of model a name: the Large Geospatial Model, or LGM.[3] We are borrowing the term because it fits. An LLM reasons over language. An LGM reasons over place.
The category
| Axis | LLM | LGM |
|---|---|---|
| Input | Text: documents, prompts, conversation | Images, plus the physical cues inside them |
| Main job | Understand and produce language | Work out where something is and what is around it |
| Grounded in | What people have written down | Geometry, terrain, weather, and how things are built |
| Answers | "What is this about?" | "Where is this, and how sure are we?" |
The two are complements, not competitors. Most verification work ends up needing both.
A lot of geolocation tools are a general model with a search box on top. You upload a photo, get a guess, and do the rest yourself. An LGM starts from the other end. The location question is the whole point, not a side effect.
That is not a fringe idea anymore. In 2024 a research model called PIGEON beat one of the best professional Geoguessr players across six straight matches, from nothing but the pixels.[1] The skill of reading a place from a photo, which used to take a person years to learn, now fits in a model. Orca 1.5 is our attempt to make that useful, not just impressive.
Under the hood it is a few specialised parts working together, each doing one job well.
The net effect is a model that reasons about a place the way an analyst would: gather the clues, weigh the strong ones over the weak ones, and say how confident the answer is.
An LGM should explain why the world fits before it names where the world is.
The same core skill shows up across a few different jobs.
Take a photo with no obvious landmark and narrow it to a smaller area worth searching.
Get from the clues in an image to a shortlist of plausible blocks or neighborhoods.
Check whether the reported place fits the road, the plants, and the weather in the photo.
Test a frame against its claimed location instead of trusting the caption.
Orca 1.x got better release by release, mostly around the edges. Orca 1.5 is a change underneath. It is built around spatial reasoning from the start, instead of treating location as an extra output on top of a general vision stack.
You notice it most in the hard cases: an ordinary road, a generic building, flat light, no metadata. That is where treating location as the main job, rather than an afterthought, pays off.[2]
Visual geolocation accuracy on four independent datasets. Standard Im2GPS thresholds. All models evaluated on the same images — no cherry-picking. The benchmark code is open source and fully reproducible.
Im2GPS3k
n=2,123 scored of 2,997 · D2 · 2026-05-31
| Model | 1 km | 25 km | 200 km | 750 km | 2500 km |
|---|---|---|---|---|---|
| Orca 1.5Oceanir | 32.2% | 64.3% | 75.7% | 86.1% | 93.3% |
| REVERSEPaper | 22.5% | 48.3% | 59.3% | 73.5% | 84.8% |
| Geo-RPaper | 18.1% | 41.5% | 58.3% | 75.3% | 86.4% |
| GAEA-7BWACV'26 | 8.4% | 36.9% | 56.0% | 73.2% | 85.6% |
| GeoCLIPNeurIPS'23 | 14.1% | 34.5% | 50.6% | 69.7% | 83.8% |
| GeoVistaAgentic | — | — | — | — | — |
| GeoAgentCVPR'26 | — | — | — | — | — |
Drop an image into the app and see where Orca 1.5 lands.
Newsletter
New cities, model releases, and what we're learning from reading places. Occasional, no noise.