Liquid AI's d1.2 and d1.2-ultra are now available on AI Gateway
- Source
- vercel.com
- Author
- Jerilyn Zheng
- Date

Context
Developers can now call Liquid AI's d1.2 and d1.2-ultra through Vercel's AI Gateway, with support for audio input and for locating objects in images (Vercel). A decision model does not write prose. It answers typed questions about an input and returns structured results: a Boolean question gives the probability that the answer is yes, a Choice question picks a category, and a Score question rates the input against a rubric. Several questions can share one input in a single request. Vercel's example hands d1.2-ultra a voicemail recording and gets back the probability that the caller is asking for a refund.
The gateway already offered decision models, including Liquid's earlier d1 and d1-ultra, which could answer questions about images, and Microsoft AI's Decision-1.1, all called through the same experimental decide function in Vercel's AI SDK. The d1.2 models add audio files as input and box questions, which locate something in an image and return its top and bottom edges as fractions of the image height and its left and right edges as fractions of the width. Vercel describes d1.2 as the faster, lower-cost model and d1.2-ultra as Liquid AI's most accurate decision model. Vercel's audio example needs version 7.0.136 or later of the ai package, and the requests show up in the gateway's logs and count toward its spending budgets.
The case for accuracy comes from Liquid's own launch, which claimed first place on the Artificial Analysis Decision Index with a score of 75. Artificial Analysis's own post confirms d1.2-ultra in first place at 75 and d1.2 at 69, each a few points above the earlier Liquid model it updates. Vercel's changelog does not list prices.
- Liquid AI's d1.2 and d1.2-ultra are decision models: instead of generating text they return structured answers to Boolean (yes/no probability), Choice (a category) or Score (a rubric rating) questions. Several questions can share one input in a single request.
- On AI Gateway the models now accept audio input. Vercel's example passes a voicemail FLAC file and reads back the probability that the caller is asking for a refund; it requires the ai package 7.0.136 or later.
- New Box questions locate an object in an image through the Decisions API, returning top and bottom as fractions of the image height and left and right as fractions of its width.
- d1.2 is the faster, lower-cost model and d1.2-ultra is Liquid AI's most accurate decision model. Decision requests appear in AI Gateway logs and count toward budgets.
- structured output — Forcing a model's response to match a schema, so downstream code can parse it instead of guessing at prose.
Checking sign-in…
Loading comments…






