Unlimited OCR Cuts Thick-Document KV Cache from 47GB to 1.77GB
xiaohu · x · 2026-10-27
Follow-up from the Wenxin Unlimited OCR Gen 2 thread: the model splits thick documents into parts, keeping only what's needed for the current section and releasing stale caches — cutting this cache's VRAM usage from 47GB to 1.77GB in tests.
Bonus: the name nods to Fate/stay night's "Unlimited Blade Works". After Gen 1, a community member rewrote its famous incantation for OCR ("pixels for body, tokens for blood..."), which the team loved enough to put at the start of the Gen 2 tech report.
Related event: Unlimited OCR Gen 2 Runs 5.6x Faster on Long Documents with Far Less VRAM(4 posts)→
More from Multimodal
- MPI's Alien Recombination method finds concept blends beyond human cognitive reach — profjamesevans · 2026-10-27
- Kling 4 generates 30-second drone-style footage, full prompt included — TheLocalLab · 2026-10-27
- Open-source ComfyUI node uses Qwen3-VL to auto-build MOSS-Image layout prompts — Maximus989989 · 2026-10-27
- Creator has Claude edit her iPhone Duo review inside Tella's agent video editor — evielync · 2026-10-27
- Qwen-Image-PE: two-stage training lifts text-to-image prompts by 7.81 points — Qwen · 2026-10-27
- Curated list of 20 open-source motion libraries for building an AI motion studio — nikola_mr64990 · 2026-10-27