Tencent's EVIE turns a scanned page into something you can actually search
Two releases land in the same place today: Tencent putting open-weight visual document retrieval models on Hugging Face, and vLLM publishing what happens when you run speculative decoding on AMD silicon instead of Nvidia's. One is about what AI can read; the other is about how cheaply you can run it.
Tencent has released EVIE-8B and EVIE-4.5B, open-weight models built to retrieve pages of documents by what they look like — and the smaller one nearly matches the big one. The pair scored 66.75 and 66.02 nDCG@10 on ViDoRe V3, the benchmark that decides whether a retrieval model can find the right page inside slides, scans, charts, and tables. Tencent validated them across 138 tasks spanning four benchmark families. The 8B model is the teacher and stores a separate 4,096-dimension vector for every token on the page rather than collapsing a document into one embedding — that is what preserves layout, typography, and table structure. The 4.5B model is distilled from it into a single compact vector, which is the version you would actually deploy: nearly the same accuracy at roughly half the parameters.
The reason this matters is that most enterprise retrieval still throws away the page. The standard pipeline runs OCR, dumps text into a chunker, and hopes the table survives — it usually doesn't. Native multimodal retrieval keeps the visual signal, and Tencent shipping a distilled 4.5B version is the part that changes the economics: a model that fits on a single GPU and can be self-hosted, not rented per page from an API. Worth watching how the 4.5B's compressed vectors hold up on genuinely messy real-world scans, since benchmark pages are cleaner than anything in a corporate archive.
vLLM measured five speculative decoding methods on AMD's MI300X and MI355X GPUs, and the headline number is that the best configurations roughly double throughput. The vLLM team, working with AMD and Embedded LLM, tested native MTP, Gemma 4 MTP, EAGLE-3, DFlash, and DSpark across Gemma, Qwen, MiniMax, and Kimi models. The top results were 2.87× for DFlash on Gemma 4 26B, 2.83× for Gemma 4's own MTP, and 2.68× for DFlash on Kimi-K2.5. Speculative decoding works by having a small draft model guess several tokens ahead, which the main model then checks in a single pass — when the guesses are right, you commit multiple tokens for the price of one forward pass.
Two caveats from the data are more useful than the peak numbers. First, the gains are wildly uneven: on Qwen3-8B, DSpark ranged from 1.15× to 1.63× depending on the workload, and some settings finished below the non-speculative baseline. Second, guessing more tokens ahead does not keep helping — throughput typically plateaued after a handful of draft tokens, which is the opposite of the "more speculation is better" intuition. For anyone serving open weights, the takeaway is that this is worth tuning per workload, not flipping on once. It also fits a pattern we noted this morning in iFlytek ships Spark X2.5, trained end-to-end on Chinese silicon — the hardware assumption under the AI stack is no longer Nvidia by default.
Huawei's Mate XT 2 borrows Samsung's privacy display, the feature that narrows the screen's viewing angle so people beside you can't read it. The Verge reports the new trifold also adopts Samsung's inward-folding design, copying a feature Samsung had reserved for the Galaxy S26 Ultra. It is a small hardware detail with a real one: on-device AI that reads what's on your screen makes shoulder-surfing a data problem, not just an etiquette one. Huawei's original Mate XT launched in China in late 2024 and internationally a few months later; Samsung has since discontinued its own trifold.
What to watch: whether EVIE-4.5B's compressed vectors hold their accuracy outside benchmark-clean documents.
If a 4.5B model can retrieve from scanned pages as well as an 8B one, does the OCR step in your stack still earn its place? Tell us in the comments.
