Tencent's EVIE, released on Hugging Face, tops the ViDoRe benchmark with 128D embeddings. The model's compact vector size slashes storage and compute, challenging the assumption that bigger embeddings mean better retrieval.
Key facts
- EVIE uses 128-dimensional embeddings
- Achieves SOTA on ViDoRe benchmark
- Released on Hugging Face by Tencent
- Model card lacks exact ViDoRe scores
- No technical paper published yet
Tencent has released EVIE, a visual document retrieval model, on Hugging Face @HuggingPapers. The headline number is the embedding dimension: 128. Most dense retrieval models ship with 768 or higher—ColBERT uses 128 per token, but EVIE compresses an entire document into a single 128D vector. That's roughly a 6x reduction in storage per vector compared to a 768D baseline, and a proportionate cut in retrieval latency when scanning millions of documents.
Why 128D matters
ViDoRe is a retrieval benchmark built for visually rich documents—scanned PDFs, forms, tables—where OCR and layout matter. The ViDoRe dataset was introduced in 2023 by the ColPali team at Illuin Technology, pairing document images with query-text pairs. Prior state-of-the-art models, like ColPali and its successors, rely on late interaction with per-patch embeddings, often running to hundreds of dimensions per patch. EVIE's single-vector 128D approach trades some expressiveness for a massive efficiency win: a 100GB document corpus indexed at 768D shrinks to ~17GB at 128D, before quantization.
Tencent hasn't published the full technical report or the exact ViDoRe scores—the Hugging Face model card lists the benchmark result but omits the mAP@5 or NDCG@5 numbers. That's a gap. Saying "SOTA" without a number invites skepticism; the community will need the leaderboard entry to verify. Still, the architectural choice is notable: single-vector dense retrieval is a known trade-off, and 128D is aggressive. For comparison, OpenAI's text-embedding-3-small uses 1536D by default, and even the compact "small" variant is 512D.
The efficiency angle
EVIE's pitch is operational. Visual document retrieval is used in enterprise search, RAG pipelines, and compliance systems, where the corpus is large and the queries are frequent. A 128D index means you can hold millions of vectors in RAM on a single machine, avoiding the need for a dedicated vector database cluster. That's a cost story, not just a benchmark story. Tencent's move signals that the next frontier in retrieval isn't accuracy alone—it's the cost per query at scale.
There's a parallel with the recent trend toward smaller, task-specific models. Google's Gemma 3 announced in March 2025 showed that a 4B model can beat larger predecessors on several benchmarks, and Mistral's 7B release in September 2023 demonstrated that efficiency can be a differentiator. EVIE applies the same logic to embeddings: don't add dimensions, cut them. The risk is that 128D is too lossy for fine-grained tasks like matching a specific table cell or a signature. The ViDoRe SOTA claim suggests it works for the benchmark's scope, but real-world documents are messier.
What's missing
Tencent did not disclose the training data, model architecture (likely a vision-language transformer), or the compute used. The model card is thin. For engineers evaluating EVIE, the absence of a technical paper is a blocker—you can't reproduce the training setup or ablate the components. The Hugging Face release is a model dump, not a research contribution. That's fine for deployment, but it limits trust. The ColPali paper from July 2024, which introduced ViDoRe, included full ablations; EVIE's release doesn't.
Still, the practical upside is clear. If EVIE holds up on internal benchmarks, it's a drop-in replacement for heavier models in document-heavy RAG systems. The 128D vector size means faster brute-force search, cheaper ANN index builds, and lower storage costs. For startups running on tight margins, that's a meaningful lever. For enterprises with millions of scanned invoices, it could be the difference between a GPU-backed retrieval service and a CPU-only one.
The release date is tied to the post's timestamp, which shows early 2026. The model is live on Hugging Face, so it's testable now. The community will likely run it against the ViDoRe leaderboard independently—that's the next signal.
Bottom line
EVIE is a compact, efficient retrieval model that claims SOTA on ViDoRe with 128D embeddings. The lack of public scores and a technical paper is a caveat, but the architectural direction is clear: smaller vectors, lower cost, and a focus on operational deployment. Tencent's move could push other labs to reconsider the embedding-size arms race.
Key Takeaways
- Tencent's EVIE model achieves SOTA on ViDoRe with 128D embeddings, a major efficiency gain.
- The release lacks technical details, but the compact vector size signals a shift toward cost-efficient retrieval.
What to watch
Watch for independent ViDoRe leaderboard runs from the community, and whether Tencent publishes a technical paper with full ablations. If EVIE's scores hold under scrutiny, expect other labs to shrink embedding dimensions; if they don't, treat the SOTA claim as marketing.
