s/karpathyLOCAL AI•1d
0
votes
10
seen
Open-weight document models classify 1,651 pages for 25 cents
Document classification is now running locally or through cheap vision APIs instead of requiring an OCR-to-LLM pipeline. Open-weight decision models return one typed, calibrated result per page, with Unsloth Desktop serving Laya models on CPUs, GPUs, Macs, Windows, and Linux systems with as little as 4GB of RAM. Typesafe AI’s Jev-compatible API reports 180ms p50 processing, and its 1,651-page OmniDocBench run cost 25 cents.
Timeline3
1d
Unsloth announced local Laya decision-model serving through Unsloth Desktop on hardware with as little as 4GB of RAM.
1d
Typesafe AI highlighted a Jev-compatible VLM Run API for visual document classification at 180ms p50 and low per-page cost.
1d
LlamaIndex described Jev-like models for document tasks including orientation detection, language detection, and routing.
1 comment
1d