This week Tencent released EVIE-Preview-4.5B on the Hugging Face open-source community: it converts each document page into a "searchable mathematical fingerprint" (the industry calls these embedding vectors), compressing storage to between 1/8 and 1/32 of comparable products — while retrieval accuracy is actually higher. On the ViDoRe V3 benchmark, this 4.5B-parameter model beat 8B competitors, taking six out of seven first-place finishes in public domains. We're more interested in the signal behind it: Chinese AI labs are no longer just chasing chatbots — they're now competing head-on in RAG (the underlying technology that lets AI find answers inside your documents), an unglamorous but critical foundational layer of the global race.
What this is
Most "knowledge" inside enterprises isn't clean Word files — it's scanned PDFs, contract images, financial reports with tables and charts. For AI to answer "what did last year's contract say about this clause," it first has to convert these documents into searchable mathematical vectors and load them into a database. Traditional approaches use 2,560 to 4,096 dimensions per page; tens of thousands of pages means tens of GB of memory. EVIE uses ColBERT-style "late interaction" with 128-dimensional vectors, compressing index size to between 1/8 and 1/32, while accuracy beats larger-parameter competitors. The base is Alibaba Tongyi's Qwen3.5-4B, fully open-source, and it drops directly into the ColPali open-source ecosystem for deployment.
Industry view
Positive views cluster around cost structure. We've noticed that teams building enterprise document AI don't lose sleep over model capability — they lose sleep over storage and indexing costs. For a mid-sized law firm, indexing 50,000 contracts can run into seven-figure RMB annually. Compact-vector schemes are the key technology for putting enterprise AI inside small and mid-sized business budgets.
But the negatives deserve air time. First, this is a Preview release — production stability at the hundred-million-document scale is unverified. Second, "cutting to 128 dimensions" is contested in the industry: some argue fine-grained retrieval must preserve high dimensionality, otherwise small details in long documents get lost. Third, Tencent open-sourced only the model weights — not the training data or evaluation pipeline — so generalization capability is in question. One more layer of observation: a flourishing garden of open-source RAG components isn't necessarily good news for enterprise IT. The integration cost of evaluating and choosing between ColPali, EVIE, Jina, and Qdrant is itself a cost.
Impact on regular people
For enterprise IT: Over the next 6–12 months, deployment costs for contract retrieval, financial-report Q&A, and compliance-review AI tools may drop noticeably — but vendor-selection risk rises in lockstep.
For working professionals: Day-to-day AI assistants for browsing PDFs and reading financial reports will continue to improve in speed and accuracy. The most direct beneficiaries are document-intensive roles: legal, finance, consulting, and bid sales.
For consumer market: Your phone's chatbot won't change because of this. It's a ToB technology with no direct connection to consumer products like ChatGPT or ERNIE Bot.