Jina AI Releases jina-ocr-v1: A 3.4B MoE Document Parser With Built-In Speculative Decoding for Low-Budget GPUs

- Jina AI dropped jina-ocr-v1 — a 3.4B MoE parser that eats PDFs, tables, and invoices and spits out Markdown at 2.57 pages/sec on one A100. Tiny 570M active params, speculative drafting head baked into the checkpoint, lossless speedup like a PoD seal that actually seals the manifest. Roast of the day: it doesn't top the accuracy charts — PaddleOCR-VL beats it 96 to 91 — so this is a hustle story, not a genius story. And the CC BY-NC 4.0 license? Free until you make money, then you owe Jina. Aggressive passive income, professionally laundered through licensing. Where's my cut? Still: lossless 1.95x speedup on an L4 is real engineering, unlike the 12,000th "revolutionary" L2 that dies in a relay window. Read the manifest before you buy the hype — speed you can use, accuracy you can't inflate.