Liquid AI's d1 decision model arrives on Vercel's AI Gateway with vision support

Vercel has added Liquid d1, a decision model from Liquid AI, to AI Gateway, its unified access layer for hosted models. According to a Vercel blog post published on October 9, developers can call the model under the identifier liquid/d1 through the same gateway they already use for other providers.

What a decision model does

Most models available through gateways today are generative: given a prompt, they produce text. d1 takes a different approach. As Vercel describes it, the model examines an application's shared state and answers typed questions built around three jobs — classifying items, routing requests, and scoring values. Rather than composing a reply, it returns a structured result with probabilities attached, and it skips text token generation entirely.

The practical effect is that a developer asking whether a support agent issued a refund — one of Vercel's own examples — gets a machine-readable answer rather than a sentence that has to be parsed before anything can act on it.

Vision input

d1 also accepts images. Vercel says the model can respond to typed questions about pictures, which points it at tasks like classifying, inspecting, and scoring visual content. The announcement demonstrates this with a local PNG file, asking the model to identify which color the image shows.

Four ways to call it

Vercel exposes d1 through several interfaces:

• the decision API in the AI SDK
• the plain HTTP API
• an OpenAI-compatible Decisions API
• a TypeSafe-compatible API

The documentation describes three question shapes — Choice, Score, and Boolean — and d1 is not the only decision model listed on AI Gateway, which suggests Vercel is treating decisions as a category in their own right rather than a one-off addition.

Logging and budgets

Because d1 runs through AI Gateway, it inherits the gateway's observability features. Vercel notes that decision calls show up in the logs and count toward budgets in the same way as any other model request, so teams can track d1 usage and spend under the controls they already have for text-generation models.

Why it matters

A large share of machine-learning work in production is not writing prose. It is routing a ticket, picking a category, grading a response, or deciding whether a step succeeded. Today that work is usually done by prompting a chat model and parsing its output, which is brittle: formats drift, models hedge, and token-level likelihoods are a poor stand-in for calibrated confidence.

A model built to return structured, probabilistic answers to typed questions removes the parsing layer. The output slots directly into ordinary code — a conditional, a routing table, a threshold check — and the attached probabilities make it possible to handle uncertainty explicitly, for instance by escalating low-confidence decisions to a person or a larger model.

Hosting d1 on AI Gateway also lowers the barrier to trying it: existing gateway users get unified access, logging, and budgeting without a new vendor relationship or a separate billing pipeline. Vision support extends the same pattern to image inspection tasks, and the presence of other decision models on the platform hints at a broader split between models that generate content and models that make calls — a distinction that could shape how applications are architected if it catches on.