A perception engine plus a research recipe for thermal, explicit about what is proven today versus what a larger GPU training run would still need to confirm.
Closed source, inquireThermal perception ships in two halves. The first is a working perception engine that takes any detector's boxes, tracks, and classes as JSON and returns thermal geometry, zero-training hotspot detection, track kinematics, and plain-language explanations, with no GPU and no training required.
The second is a research half: a radiometric data-synthesis engine, sensor-harmonization tooling, a self-supervised pretraining harness, and frozen evaluation protocols, built before any model exists, so a future GPU run plugs a model in and reads numbers out instead of grading its own homework.
The project's README is direct about the boundary: there is no checkpoint in the repo and no claim that a production thermal foundation model exists yet.
Hotspot detection precision 1.0, recall 1.0 (4 of 4 true positives, 0 false positives out of 4 ground-truth objects); an on-axis metric round trip exact to the millimeter (0.500 m wide by 1.750 m tall at the true position); a walker track measured at 1.062 m/s against a true 1.0 m/s.
1x RTX 5060 Ti, 4096 synthetic frames deduplicated to 2503, a ViT-S/16 encoder (about 24.7M parameters), batch 256, loss descended 0.92 to 0.69 over 9,815 steps and 32.0 minutes, throughput about 1,306 images per second, peak VRAM 4,875 MB.
The same trainer re-pointed at LLVIP, a real, public, non-commercial-license long-wave infrared pedestrian dataset (ICCV 2021 Workshops), 15,488 real infrared JPEGs. Loss descended 0.98 to 0.65 (about 33 percent) over 12,974 steps and 40.0 minutes, throughput about 1,382 images per second, peak VRAM 4,874.6 MB, essentially identical to the synthetic pilot since only the data source changed. Both corpora sit 13 to 80 times below the project's own documented 200,000-frame overfitting floor, so neither pilot is evidence of learned-feature quality; both are evidence that the trainer, corpus builder, and math work end to end on genuine sensor data.
A gated Hugging Face model, thermal-mae-vits-llvip-pilot, exists with a real checkpoint file (about 85.6 MB total repo storage), tagged as a training pilot. Its own model card could not be read this session because Hugging Face's access gate restricted it; only public API metadata (file listing, tags) was confirmed.
An outsider can try the demo Space and download the published benchmark dataset to see the synthetic hotspot-detection engine behavior. The gated model requires Hugging Face access approval, which was not granted this session. The MAE training-mechanics pilots are not independently reproducible without both the private source and the LLVIP dataset's own non-commercial license terms.