Refraction
Pl. IVision models that bend toward detail. A hairline crack in a turbine blade, a shadow on an X‑ray, a single misprinted digit on an invoice.
PRISMA Labs builds perception models that take in images, sound and text the way a prism takes in light: everything passes through, every part is separated, nothing is lost.
Vision models that bend toward detail. A hairline crack in a turbine blade, a shadow on an X‑ray, a single misprinted digit on an invoice.
Language models that split a dense document into its parts: claims, numbers, obligations, tone. Each part is shown on its own line, with the source beside it.
Multimodal fusion. When a camera, a microphone and a report disagree, PRISMA shows where the waves meet and where they cancel out.
| Model | Sees | Band | Context | Latency | Status |
|---|---|---|---|---|---|
| Clear 2 | Images, video | 4,096 frames | 38 ms | Available | |
| Lux | Documents, text | 2,000,000 tokens | 120 ms | Available | |
| Halo | Sound, speech | 9 h audio | 64 ms | Available | |
| Spectrum | All of the above | Mixed | 210 ms | Preview |
Fine detail recognition on the Lumen 40K benchmark, 2026 edition. Higher is better. Every run is published with its settings, so anyone can repeat it.
Sample figures for this concept site
Light can bend without being lost.
光は、曲がっても失われない。