alphaXiv launches quickarXiv — auto-generated blog explainers for arXiv papers powered by GLM OCR

TL;DR

alphaXiv is shipping a URL-swap feature: replace arxiv.org with quickarxiv.org on any paper link and get an auto-generated blog-style explainer with extracted figures, key insights, associated code, and the author’s launch tweet. The extraction pipeline is now powered by Z.ai’s GLM OCR (the same lineage as the GLM-5.x vision family already filed on the wiki). The demo target is @a1zhang’s Prime Agent paper (2608.23552). This is a product announcement, not a research result — no eval numbers, no methodology beyond “we switched OCR backends.”

Key claims

• URL rewrite arxiv.org → quickarxiv.org renders an auto-generated explainer page with figures, explanations, code links, and the author’s tweet [tweet body].
• The extraction backend has moved to Z.ai’s GLM OCR from an unnamed prior stack [tweet body].
• Working example: quickarxiv.org/abs/2608.23552 on Prime Agent [reply tweet 2092627954400682456].

Method

Not disclosed. Announcement text names GLM OCR as the extraction model but does not describe the summarization/explainer generation pipeline, layout parsing, or how figures are extracted and re-anchored to text. Presumably: OCR the PDF with GLM OCR → run a downstream summarizer / explainer LM → surface extracted figures inline → cross-link the author’s launch tweet and code repo when available.

Results

None reported. 62.2K impressions on the launch tweet at time of filing; no benchmark, latency, cost, or accuracy claims.

Why it’s interesting

This is a productization story rather than a research result, but it lands on a recurring wiki thread: what OCR backend do you pick when you want to bulk-process arXiv at low cost? alphaXiv’s previous public number was extracting datasets from 500K+ papers for ~$1000 using DeepSeek-OCR versus ~$7500 with Mistral OCR (alphaxiv used DeepSeek-OCR to extract datasets from 500k+ arXiv papers for $1000 (vs $7500 with Mistral OCR)) — so their choice of extraction backend is a live signal about the current cost/quality frontier. The switch to GLM OCR is notable because Z.ai’s GLM-5.x vision lineage (GLM-5.3-Flash — 320B/18B natively multimodal hybrid sparse+linear attention MoE with mHC, GLM-5.2-Vision-NVFP4 — bolting MoonViT onto frozen GLM-5.2 with a 49.5M projector) has been quietly competitive with DeepSeek on multimodal benchmarks; and quickarXiv’s summarizer output is the kind of dense-figure-plus-text artifact that stresses the “render text as pixels” thesis Karpathy pushed off DeepSeek-OCR (Karpathy: pixels may be better LLM inputs than text tokens (DeepSeek-OCR springboard)).

See also

• alphaxiv used DeepSeek-OCR to extract datasets from 500k+ arXiv papers for $1000 (vs $7500 with Mistral OCR) — alphaXiv’s earlier DeepSeek-OCR vs Mistral OCR cost comparison; sets the baseline this backend swap is measured against.
• DeepSeek-OCR: Contexts Optical Compression — the model quickarXiv appears to be replacing under the hood.
• DeepSeek-OCR 2: Visual Causal Flow — DeepSeek-OCR 2 follow-up in the same lineage.
• GLM-5.3-Flash — 320B/18B natively multimodal hybrid sparse+linear attention MoE with mHC — Z.ai’s current GLM-5.x flagship; provides the OCR model family now doing the extraction.
• MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing — comparable document-parsing VLM from Shanghai AI Lab.
• PaddleOCR-VL: Boosting Multilingual Document Parsing via a 0.9B Ultra-Compact Vision-Language Model — another competitive open OCR VLM in the same weight class.