Liquid AI Launches d1-ultra Decision Model on Vercel AI Gateway

Liquid AI and Vercel announced the general availability of d1-ultra on Vercel AI Gateway, introducing a specialized multimodal decision model engineered specifically for high-throughput enterprise classification, routing, and scoring across text and image workloads. Rather than generating open-ended conversational text that requires fragile post-processing and regex parsing, d1-ultra operates as a deterministic decision engine that outputs structured categorical selections paired with rigorously calibrated probability distributions.

The launch arrives as enterprise engineering teams increasingly encounter the latency, cost, and hallucination penalties of routing mission-critical business logic through generative autoregressive models. By restricting the model architecture to classification and decision-theoretic inference, Liquid AI delivers sub-second deterministic evaluation for visual inspection, automated content moderation, document validation, and complex customer claim verification workflows across high-traffic digital platforms. The system enables companies to automate manual inspection steps while substantially slashing computational overhead.

Multimodal decision routing and scoring

The core capability of d1-ultra is its ability to ingest dense visual evidence alongside complex prompt constraints to reach decisive routing conclusions. In insurance claim automation, for example, enterprises can submit photographic vehicle damage evidence together with policy documentation to automatically verify damage severity and triage claims without human intervention.

Similarly, in manufacturing quality control, high-resolution assembly line captures are compared against defect taxonomy datasets in real time. Because the architecture produces native calibrated confidence scores alongside categorical assignments, backend pipelines can establish automated routing thresholds—automatically clearing high-confidence items while routing edge cases directly to specialized human review queues.

Calibrated probabilities replace generative hallucination

Traditional large language models struggle with statistical calibration: when prompted with classification tasks, their token probability distributions often reflect extreme overconfidence or sensitive prompt phrasing biases. The d1-ultra architecture was trained specifically to output mathematically grounded, well-calibrated posterior probabilities that reflect empirical accuracy across real-world operational distributions.

This calibration allows infrastructure architects to treat model outputs as true probabilistic priors. Software teams can implement programmatic decision gates where actions requiring high precision—such as financial transaction flagging or access permission revocation—trigger only when the model probability exceeds strict regulatory thresholds, dramatically reducing false positives compared to standard prompt-engineered frontier models.

Three structured query modes: Boolean, Choice, and Score

Developers interact with d1-ultra through three native structured evaluation primitives that eliminate the need for schema parsing frameworks like Zod or JSON repair libraries. The Boolean mode evaluates binary truth claims, returning normalized yes-or-no probability pairs. The Choice mode accepts an arbitrary list of named categorical labels and calculates exact softmax-normalized probabilities across all candidate options.

The Score mode evaluates inputs against ordered numerical scales, such as rating customer satisfaction sentiment from 1 to 5 or assessing code quality severity levels. Multiple independent queries can be batched against the same multimodal payload within a single HTTP round-trip, allowing systems to concurrently check compliance, safety, and priority in parallel without incurring multiple sequential network requests.

32k context window and zero output token pricing

Architecturally, d1-ultra features an expansive 32,768-token context window that comfortably accommodates lengthy technical documents, multi-page PDF renders, and multiple high-resolution images within a single evaluation cycle. This capacity ensures that complex context documents and reference schemas can be evaluated alongside visual inputs without truncation or multi-pass chunking strategies.

In a significant departure from standard foundation model pricing structures, Vercel and Liquid AI have implemented a billing model that charges exclusively for input tokens—including image inputs—with zero billing for output tokens. Because classification decisions output concise structured records rather than hundreds of generative output tokens, operational inference expenses scale predictably with workload volume rather than output verbosity. Enterprise users report substantial cost reductions on continuous document review tasks, transforming formerly expensive batch processes into continuous, real-time data flows.

Production integration via Vercel AI SDK

Integration into existing production backends is supported natively across multiple standard interfaces. Teams using the Vercel AI SDK can reference the model directly using the identifier liquid/d1-ultra, gaining immediate access to built-in retry handling, caching layers, and telemetry tracing across edge infrastructure.

For multi-cloud or non-Node environments, the service exposes an OpenAI-compatible Decisions endpoint as well as a TypeSafe REST API. Engineers can drop d1-ultra into existing proxy routers without modifying transport payloads, making it possible to migrate brittle prompt-based classifiers to dedicated decision infrastructure in an afternoon while maintaining strict latency and reliability guarantees.