EVIE 8B released
- Source
- huggingface.co
- Author
- Tencent model releases
- Date
Tencent released EVIE-8B, a late-interaction visual document retrieval model built at 4096 dimensions that now tops the ViDoRe V3 leaderboard. The model reads document images directly rather than relying on OCR text extraction, which helps it handle tables, charts, and mixed layouts that trip up text-only retrieval systems. Its main practical role in Tencent's lineup is as the distillation teacher for EVIE-4.5B, the smaller, compressible model released alongside it, meaning EVIE-8B sets the accuracy ceiling that the compressed variant is measured against. For developers who need the highest possible retrieval quality on complex documents and can afford the larger index size, EVIE-8B is now the strongest open option on this benchmark. Teams optimizing for storage instead should look at the distilled EVIE-4.5B released the same day.
Context
Tencent's EVIE-8B retrieves document pages by comparing a written query with images of those pages. It represents the query and page as multiple vectors, then scores relevance by matching each query component to its strongest counterpart in the document. This gives the retrieval process access to visual structure, including tables and charts, alongside text. The model card provides code that produces relevance scores from a query and document image, making the model a component for finding supporting pages.
- distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
Checking sign-in…
Loading comments…