TopK Releases topk-embed-v1 to Slash Visual Document Search Storage Costs
TopK released an open-source family of multimodal multi-vector embedding models with Matryoshka representations, 1-bit compressibility, and a hosted tier at five cents per million tokens.
- TopK released topk-embed-v1, open-source multimodal multi-vector embedding models for text and page images.
- Two sizes under Apache 2.0: xsmall at 0.8B and small at 2B, both on a Qwen3.5 hybrid backbone.
- Hosted production tier priced at $0.05 per 1M tokens inside the TopK retrieval engine.
- Matryoshka embeddings truncate from 2048 down to 64 dims, compressible to 1 bit per dimension.
- On ViDoRe v3 image retrieval the small model averages 65.22 nDCG@10, 64.91 at 256 dims.
- Uses MaxSim late-interaction scoring and integrates with sentence-transformers MultiVectorEncoder.
TopK cuts the storage cost of late-interaction retrieval
TopK has released topk-embed-v1, a family of multimodal retrieval models that searches text and page images with token-level embeddings. The company published two open-weight checkpoints under Apache 2.0 and listed a hosted production model at $0.05 per 1 million tokens.
The release targets the main operational drawback of late-interaction retrieval: storing many vectors for every document. Matryoshka truncation, 1-bit quantization, and token pooling reduce that footprint while preserving most of the reported retrieval quality.
| Offering | Size | Availability |
|---|---|---|
topk-embed-v1-xsmall |
0.8B parameters | Open weights, Apache 2.0 |
topk-embed-v1-small |
2B parameters | Open weights, Apache 2.0 |
| Hosted production model | Undisclosed | TopK service, $0.05 per 1M tokens |
Late interaction preserves document detail
A conventional dense retriever compresses each document into one vector. That representation can lose rare names, code identifiers, secondary topics, and details confined to one region of a scanned page.
Multi-vector retrievers retain one vector per text token or image patch. During MaxSim scoring, every query vector selects its closest document vector, then the system sums those matches into one relevance score. The method preserves fine-grained evidence, but it requires more storage and a retrieval engine capable of matrix-based scoring.
At publication, 18 of the top 20 systems on the public ViDoRe v3 visual-document leaderboard used multi-vector late interaction. A 4.80B late-interaction model also exceeded the best 8.14B dense model by 3.76 leaderboard points, according to the project’s model card.
Three controls shrink the index
Topk-embed-v1 reduces multi-vector storage through three configurable techniques:
- Matryoshka truncation: The models produce 2,048-dimensional embeddings whose prefixes can be truncated to 1,024, 512, 256, 128, or 64 dimensions. Developers can select 256 dimensions with
config_kwargs={"output_dim": 256}. - One-bit quantization: Training prepares the representations for compression to one bit per dimension, reducing the bytes required for each retained vector.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
