Liquid AI open-sources d1: on-device decision models in 8ms
Weights for text, vision and audio decisions you run on a Jetson, an open Arabic-first ASR with timestamps, and new Vercel Pro storage fees from Oct 23.
Copy markdown
Decision models you download, not rent
Liquid AI open-sourced d1-3B (text + images) and an experimental d1-omni-600M (which adds audio): both return a typed answer with probabilities in a single forward pass — zero decoded tokens — at 8ms for a yes/no on an RTX 4090 or 50ms on a Jetson Orin Nano, with day-one llama.cpp support. d1-3B scores 48.57 on the Decision Index, matching a 35B hosted model, so spam filters, request routers and urgency scorers can run on your own edge hardware instead of a paid API.
An open multilingual ASR that nails Arabic
TII's Falcon-ASR is a 1.6B open speech model that transcribes Arabic (including the Emirati dialect), English, French, Spanish and Portuguese from one set of weights — no language flag — with word-level timestamps for alignment. It posts 5.74% WER on English and 20.92% on Arabic, ahead of the prior leaderboard best, giving multilingual transcription pipelines a small model with strong Arabic and built-in timestamps; a Hugging Face demo space is live now.
Vercel starts billing Pro deployment storage
Existing Pro teams now pay $0.10/GB-month for Deployment and Functions Storage, and the retention window drops to 30 days on Oct 23 (new Pro teams already default to 30). Prune or export older deployments you want to keep before the cutover.