VultronRetriever targets cheaper, offline search for AI agents
The VultronRetriever model family has been released on . It was announced at Raise Summit Paris, with a demo showing question answering and document embedding running fully offline on an iPhone. Each model is presented as the top model in its size class on the , and VultronRetrieverPrime-8B is presented as the overall number one model.
Prime-8B is said to use up to 16 times less index storage and deliver up to 12 times higher throughput than earlier leading 9B-class models. VultronRetrieverCore-4.5B ranks just behind Prime and is said to beat models twice its size. VultronRetrieverFlash-0.8B is said to outperform models up to five times larger, run cool on edge devices, and index up to 60 images per minute while fully offline.
With Hydra , the models are said to support with very high precision, while generation can use up to half the memory of comparable models. The is described as having no cross-dataset duplication and no evaluation contamination, with no overfitting seen in private MTEB tests.
Key points
- The VultronRetriever model family is now available on .
- An offline iPhone demo showed question answering and document embedding.
- Prime-8B is said to cut index storage by up to 16 times and raise throughput by up to 12 times versus earlier leading 9B-class models.
- Flash-0.8B is designed for edge devices and can index up to 60 images per minute offline.
- Hydra is said to reduce memory use for generation by up to half compared with similar models.