Plexe
Small models that each do one job on travel data: group a hotel's rooms across suppliers, or find the place a traveller means from half a word. Each is built for its task, measured before every release, and fast enough to call live.
The task: take one hotel's room names from every supplier and group the ones that are the same room. Benchmarked on 6,337 blind hotels from around the world and from live customer traffic, never used to train or tune the model, against one written definition of a room.
| Level | System | F1 | Purity | Completeness |
|---|---|---|---|---|
| Same room | Leading provider | 82.3 | 83.7 | 80.9 |
| Same room | mlr-1 | 91.7 | 94.5 | 89.1 |
| Room type | Leading provider | 61.3 | 93.7 | 45.6 |
| Room type | mlr-1 | 95.1 | 96.0 | 94.1 |
B³ F1 over every supplier name. Purity: each card holds one room. Completeness: each room sits on one card. Labels follow our published room definition, made by two independent LLM referees with an adjudicator. Both systems received the same supplier room names. 81.6% of hotels resolve in under a second, timed end to end by the caller. 16 Nov 2026.
GPT-5.5 and GPT-6.1 match mlr-1's F1 at 17+ seconds a hotel: mlr-1 is 67× faster and 44× cheaper than GPT-6.1 at the same F1.
| Model | F1 | Purity | Completeness | Seconds per hotel |
|---|---|---|---|---|
| GPT-5.5 | 94.1 | 92.1 | 96.2 | 44.0 |
| GPT-6.1 | 93.6 | 91.3 | 96.0 | 17.3 |
| mlr-1 | 93.0 | 94.2 | 91.9 | 0.26 |
| gpt-oss-120b open | 89.0 | 96.6 | 82.4 | 25.8 |
| JEV (Typesafe) | 86.2 | 92.1 | 80.9 | 0.38 |
| DeepSeek V3.2 open | 68.2 | 58.7 | 81.5 | 4.7 |
| Qwen3 32B open | 59.3 | 51.4 | 70.1 | 3.1 |
| Llama 3.3 70B open | 52.1 | 39.0 | 78.5 | 1.6 |
American Inn by OYO Colorado City I-20, Colorado City, from the benchmark, resolved in 203 ms. Hover a room to see its names.
1 in 5
| Rule engines | Decision model | |
|---|---|---|
| Built | Per supplier, by hand | Once, for every supplier |
| Taught by | Patterns and code tables | A written definition of a room |
| New wording | Breaks | Reads it |
| New customer | Rebuilds the work | Clean rooms from the first call |
One request maps one hotel's rooms from all its suppliers. Not priced per room or supplier.
$0.20 per 1,000 requests
Custom
A request is one successful API call; in a bulk upload, each hotel is one request. Errors and playground calls are free. Prices exclude taxes.
Travel inventory reaches you from dozens of suppliers, each describing the same hotel, room and photo in its own way. Our first models resolve that, behind one API.
Groups every supplier's offers for a hotel by room, and those rooms into room types. Reads names as suppliers write them, with any beds, views and features they send.
Try it on a real hotel →Turns what a traveller types into ranked cities, areas, landmarks, airports and hotels, with the area to search. Reads half words, typos and any script.
Try the search box →An image search model for travel content. Details on release.
Resolves each supplier's hotel record to one property, from its name, address and location.