· via Vercel blog
Liquid AI's d1.2 decision models arrive on Vercel AI Gateway with audio and bounding box support
Vercel has added Liquid AI's d1.2 and d1.2-ultra decision models to AI Gateway, letting developers query audio and images for structured answers such as probabilities, categories and bounding boxes.

Vercel has added Liquid AI's d1.2 and d1.2-ultra models to its AI Gateway, and they bring a very different interface from the usual chat-style API. According to the Vercel blog, these are decision models: instead of generating text, they return structured answers about whatever input they are given.
Structured answers instead of generated text
The models are driven by typed questions rather than free-form prompts. According to Vercel, developers can ask Boolean questions, which return yes/no probabilities; Choice questions, which select a category; and Score questions, which rate an input against a rubric. Multiple questions can share the same input in a single request, so one file or piece of content can be evaluated along several dimensions in one call.
Audio input for classification
Audio support is a notable part of this release. Vercel's example reads a voicemail file, voicemail.flac, and returns the probability that the caller is requesting a refund. That is a classification task that would normally involve transcribing the audio first and then analysing the text. The example runs on AI SDK 7, requires version 7.0.136 or later of the ai package, and passes the audio as a file part in state.
Bounding boxes for images
The models also accept images. For classification, image bytes are passed along with the matching media type, such as image/png. A separate Box question type returns coordinates: according to Vercel, top and bottom come back as fractions of the image height, and left and right as fractions of its width. The company's Python example uses the Decisions API to locate a signature in a form image and print its bounding box, a job usually handled by dedicated object-detection systems.
Which model to pick, and how it is billed
The two variants split along speed and accuracy lines. According to Vercel, d1.2 is the faster, lower-cost option, while d1.2-ultra is Liquid AI's most accurate decision model. They are addressed as liquid/d1.2 and liquid/d1.2-ultra, and requests require an AI Gateway API key stored as AI_GATEWAY_API_KEY.
Operationally, decision requests appear in AI Gateway logs and count toward budgets, the same as other gateway traffic. Teams already routing model calls through Vercel can therefore add decision endpoints without new metering or observability plumbing.
Why it matters
Most model integrations are built around generating text and then extracting structure from it: prompt a chat model, ask for JSON, parse the reply, handle the edge cases where the output drifts. A decision model inverts that contract. The structure is the interface itself, whether that is a probability, a category, a rubric score or a set of coordinates, which removes a common source of fragility in classification-heavy pipelines such as triage, routing and form processing.
The audio capability matters for the same reason. Voicemail sorting, call routing and similar voice workflows can skip a separate transcription step, since the classification happens directly on the audio. Bounding-box output extends the models beyond labelling content to actually locating things within images, territory usually reserved for specialised computer-vision services.
For Vercel, adding a non-generative model family to AI Gateway suggests the gateway is positioned as a general routing layer for inference, not just a proxy for chat models. For Liquid AI, it puts its decision models in front of developers who are already building on Vercel's platform.
- #liquid-ai
- #vercel
- #ai-gateway
- #machine-learning
- #api