jina-ocr-v1: open-weight OCR turns PDFs to Markdown on an L4
The 3.4B, 570M-active parser hits 91 on OmniDocBench — plus a libheif RCE behind Slack, GitHub and Meta, and Zing-0.5, an open real-time world model.
Copy markdown
Open weights and an OpenAI-compatible endpoint
jina-ocr-v1 is a 3.4B MoE parser (~570M active) that turns PDFs, scans and images into Markdown with tables and LaTeX — 91.14 on OmniDocBench at ~2.57 pages/sec on one A100, and FastMTP speculative decoding gives ~1.95x on a budget L4. Weights are on Hugging Face (BF16, ~6.8GB, CC BY-NC 4.0) with an OpenAI-compatible API, so a local doc-to-Markdown pipeline is a genuine weekend build.
HEIF Heist: one uploaded image, remote code execution
The 'HEIF Heist' disclosure details a heap overflow in libheif/libde265 — the decoders behind Slack, GitHub Enterprise, Meta, AWS and any ImageMagick or Next.js/Astro/Gatsby image pipeline — where a single malicious HEIF/AVIF upload chains to RCE, the same bug an AI-driven team rode into OpenAI's monorepo. If your app decodes user images, update libheif to at least 1.23.4 now and sandbox or disable HEIF/AVIF wherever you don't need it.
Zing-0.5: an open world you can actually play
Seedleap open-sourced a 5B autoregressive world model — weights, inference code and an SGLang server — that renders an explorable 832x480 scene at 24 fps you steer live with keyboard and text prompts, for roughly $0.009 a stream-minute. It's a creative unlock: a promptable, playable environment to build a demo or game prototype around instead of a fixed video clip.