Liquid AI's d1.2 and d1.2-ultra are now available on AI Gateway
Liquid AI made its d1.2 and d1.2-ultra decision models available on Vercel AI Gateway with audio inputs, bounding box outputs, and structured answers.
By Dillip Chowdary • Nov 19, 2026 • Source: Vercel Blog

At a glance
- Product
- Liquid AI d1.2 and d1.2-ultra
- Availability
- Vercel AI Gateway
- SDK version
- ai 7.0.136 or later
- Input formats
- Audio flac and image png
- Question types
- Boolean, choice, score, box
- Operational limits
- Requests count toward gateway budgets
In this post
Liquid AI launched its d1.2 and d1.2-ultra decision models on Vercel AI Gateway, adding native audio processing and spatial bounding box detection to the managed platform. As detailed in Vercel Blog's report, these systems discard traditional freeform text generation in favor of deterministic, typed responses for structured classification and evaluation. Engineering teams can route production workloads through unified gateway endpoints while standardizing scoring, categorizing, and spatial localization across multimodal inputs.
This guide details the technical capabilities, architectural differences, and code integration patterns for both new models. It is designed for backend engineers and machine learning practitioners who build automated classification pipelines, content moderation checks, and multimodal verification systems. Readers will learn how to configure decision calls, structure questions across shared inputs, and process audio or image assets using standard client SDKs.
01Decision models with structured answers
Decision models represent a fundamental departure from generative language models by outputting strict structured classifications instead of streaming unstructured tokens. Rather than parsing messy conversational text with brittle regular expressions, developers query the model with concrete Boolean, Choice, Score, or Box requests. Every query evaluates against the provided input and returns validated mathematical probabilities, selected labels, or numerical ratings.
This architectural shift eliminates hallucinated tokens and token parsing overhead from critical business logic. Applications can evaluate complex multimodal files with predictable schemas, ensuring that downstream database records and automated triggers receive clean values. By returning typed outputs directly, the gateway models streamline operational workflows across customer service routing, document scanning, and fraud detection.
02Tradeoffs between d1.2 and d1.2-ultra
Liquid AI structured the release into two distinct tiers to accommodate different latency, throughput, and budgetary constraints. The base d1.2 model serves as the faster, lower-cost option designed for high-volume classification tasks where processing speed and operational efficiency take precedence. Conversely, d1.2-ultra acts as Liquid AI's most accurate decision model, providing higher precision for nuanced evaluations and edge cases.
| Model | Role and profile | Relative cost | Latency profile | Target accuracy | Model identifier |
|---|---|---|---|---|---|
| d1.2 | High-volume classification | Lower cost | Faster execution | Standard decision tier | liquid/d1.2 |
| d1.2-ultra | High-precision evaluation | Higher cost | Standard execution | Most accurate tier | liquid/d1.2-ultra |
Teams can route preliminary checks through the lightweight d1.2 tier before escalating ambiguous inputs to d1.2-ultra. Because both options share identical input specifications and query structures, swapping between them requires altering only the model string parameter. This consistency allows engineering teams to optimize runtime costs without refactoring existing data ingestion pipelines.
03Boolean choice score and box queries
The platform supports four specialized question types that can be combined across a single multimodal input. Boolean questions compute calibrated yes or no probabilities, Choice questions classify inputs into discrete categories, and Score questions rate content against defined rubrics. Furthermore, Box questions locate specific objects within images and return spatial boundaries.
Advertisement
Tech Pulse Daily
Get tomorrow's pulse first
Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.
Spatial localization through Box questions outputs normalized coordinates where top and bottom bounds represent fractions of image height, while left and right bounds represent fractions of image width. Multiple questions can evaluate the identical input within a single API request, reducing network overhead and duplicate payload transfers. For example, a single image can yield a category choice, a quality score, and a coordinate bounding box simultaneously.
04Setup steps for AI SDK and Python
Developers can interact with the models through either the TypeScript AI SDK or direct REST calls against the Decisions API endpoint. TypeScript environments require the ai library at version 7.0.136 or higher, which exposes the experimental_decide function for audio and image evaluation. Pass raw audio files as state objects with appropriate MIME types like audio/flac to evaluate vocal intent.
- Set authorization key export AI_GATEWAY_API_KEY with your active Vercel AI Gateway key credential.
- Install AI package update npm project dependencies to require ai 7.0.136 or later.
- Configure decision call pass model liquid/d1.2 or liquid/d1.2-ultra alongside input state and questions.
- Send API request submit payload to the decisions endpoint at https://ai-gateway.vercel.sh/v1/decisions.
npm install ai@7.0.136export AI_GATEWAY_API_KEY="your-ai-gateway-api-key"Python applications can invoke the Decisions API directly by sending base64-encoded file buffers within JSON payloads to the gateway endpoint. The HTTP request requires an authorization bearer header containing the gateway key and returns an answers array with detected coordinates or classification scores. This dual client support ensures compatibility across both frontend edge functions and backend data pipelines.
05Gateway logging and budget tracking controls
Deploying decision models through AI Gateway automatically hooks into unified operational monitoring and cost governance tools. Every decision request appears in centralized AI Gateway activity logs, providing visibility into execution latency, status codes, and model invocation rates. Furthermore, all executed decision calls count directly against configured organizational spending budgets to prevent accidental cost overruns.
Note: Decision requests appear in centralized AI Gateway activity logs and count toward organizational spending budgets. Developers must configure the AI_GATEWAY_API_KEY environment variable before running requests.
Teams should monitor log telemetry when rolling out audio classification and spatial object detection to verify payload sizing and latency behavior. System administrators can establish budget limits across development, staging, and production environments to maintain predictable cloud expenditures. As decision workloads expand, unified logging ensures complete observability across all Liquid AI model transactions.
Frequently asked questions
What models did Liquid AI launch on AI Gateway?
Liquid AI launched d1.2 and d1.2-ultra on Vercel AI Gateway. The d1.2 model serves as the faster, lower-cost option, while d1.2-ultra provides the highest decision accuracy.
How do Liquid AI decision models differ from generative models?
Decision models return structured answers rather than generating open-ended text. They evaluate inputs against Boolean probabilities, categorical choices, numerical scores, or spatial bounding boxes.
What minimum library version is required to classify audio with AI SDK?
Audio classification requires version 7.0.136 or later of the ai npm package. Developers pass audio files such as FLAC recordings directly in the state parameter using experimental_decide.
What coordinates do bounding box questions return?
Box questions return top and bottom coordinates as fractions of image height, alongside left and right coordinates as fractions of image width. These normalized coordinates pinpoint located objects inside images.
Sources
Author
Dillip Chowdary
Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.
Related on Tech Bytes
Advertisement