TypeSafe AI Releases Jev 1.14: Decision Model Lands on AI Gateway with Image and Long-Context Support

TypeSafe AI's decision model Jev has received its 1.14 update and is now officially available on AI Gateway. Unlike traditional large models that generate free-form text, Jev focuses on structured decision tasks — you pass in a state and a set of questions, and it returns typed Choice, Score, and Boolean answers with probabilities.

What Is a Decision Model?

Jev positions itself as a "decision model for software." In real engineering scenarios, many AI needs aren't about having a model write a paragraph — they're about having it make a judgment among a limited set of options: which category does this content belong to, which pipeline should this request be routed to, does this log entry indicate an anomaly, does a particular image meet the criteria.

The traditional approach is to wrap these questions into prompts and then parse the model's free-text response. This approach isn't just expensive to parse — it's also prone to format jitter that causes downstream failures. Jev's approach is to eliminate this uncertainty at the API level: all questions are requested in parallel at once, and typed answers with corresponding probabilities are returned directly.

According to the official description, the model covers classification, routing, scoring, and automated validation scenarios, making it better suited for embedding into existing software systems rather than existing as a standalone conversational application.

Typed Output is the most fundamental difference between Jev and traditional large models. The output of a traditional LLM is a token sequence — essentially a string. Even when constrained via JSON mode or function calling, the return value remains opaque at the type-system level — downstream code must perform runtime validation, for instance parsing with Zod or Pydantic and falling back to a degraded branch when parsing fails. Jev moves this constraint forward into the model API itself: the caller declares field names and types (boolean, choice, score) in the questions parameter, and the model returns a structured object that already satisfies that schema, with no secondary parsing required.

Probability Output addresses another engineering pain point. The "certainty" of free text can only be expressed implicitly through wording (such as "very likely" or "perhaps") and cannot be quantified. Jev attaches a 0–1 confidence score to each answer, enabling downstream code to write explicit threshold logic, such as if (answer.probability < 0.8) routeToHuman(). This turns AI decision-making from a "non-deterministic black box" into an observable, routable, auditable component — the key precondition for embedding it into software systems.

Two Key Updates: Image Input and Long Context

Images as Part of the State

Starting with Jev 1.14, image input is accepted. Developers can place an image as a file part alongside text in the state, then ask Boolean or other typed questions about the image content.

The official example reads a local PNG product photo and asks "Is the image predominantly red?" The code calls the model typesafe-ai/jev-1.14.0 via the AI SDK's experimental_decide, with the state array containing both a text description and a file part:

const result = await decide({
  model: 'typesafe-ai/jev-1.14.0',
  state: [
    { type: 'text', text: 'A product photo from a support ticket.' },
    { type: 'file', mediaType: 'image/png', data: readFileSync('./photo.png') },
  ],
  questions: {
    red: { type: 'boolean', instructions: 'Is the image predominantly red?' },
  },
});

The model returns result.answers.red as a boolean answer with probability. This pattern is particularly practical for workflows like automatic support-ticket classification and content moderation that require combining visual information with judgment.

AI SDK is an open-source toolkit from Vercel for calling models from different vendors in a unified way within TypeScript/JavaScript. Through its provider abstraction layer, the same calling code can switch between OpenAI, Anthropic, or TypeSafe AI in this example, without changing business logic.

The separation of state and questions is the core design of the Jev API. state describes "the current situation" — it can be a piece of text, an image, a segment of logs, or even a mix of multiple modalities. questions describes "what to ask" — multiple fields can be declared at once, and the model returns each answer in parallel. This separation offers two benefits: first, the same state can be reused across multiple question sets, reducing redundant passing; second, the question set itself is typed, so frontend code can statically infer the return value types from it, achieving end-to-end type safety. The experimental_ prefix in experimental_decide signals that the API is still in rapid iteration and that its signature may change.

Experimental Long-Context Mode

For documents, logs, or very long conversation histories, Jev 1.14 offers an experimental long-context mode, enabled via providerOptions.typesafe.longContext: true.

The official example reads a support log and determines "whether the log mentions that a delivered item arrived damaged." Similar to image input, developers simply pass the long text as state along with a Boolean question to get a structured judgment.

The long-context mode is still labeled "experimental," meaning stability and cost may still change — suitable for early evaluation rather than direct use in critical production paths.

Long Context refers to the maximum input length a model can process in a single request. In normal mode, the model might accept only a few thousand tokens; enabling longContext can extend that to tens of thousands or more, at the cost of higher latency and expense, and attention decay over extremely long inputs may affect judgment stability.

For decision tasks, the significance of long context is that many judgments depend on the full text rather than local clues. For example, in "whether the log mentions that a delivered item arrived damaged," the damage description might appear on line 500, while irrelevant exception stack traces occupy the first 4,000 lines. Truncating the input could cause the model to miss key evidence; doing retrieval before judgment introduces additional recall-failure risk. Long-context mode lets the entire state enter the model, reducing the number of layers in the engineering chain — but precisely because cost and stability aren't yet settled, it's officially labeled "experimental," and it's recommended to try it first in fallback-capable evaluation workflows.

Enterprise Features and Ecosystem Integration

Jev offers options clearly aimed at enterprise customers regarding compliance and data security:

• Zero Data Retention: Can be enabled per request, with a commitment not to retain data
• No Training: Can be enabled per request, with a commitment not to use data for training
• Observability: Decision calls appear in logs and custom reports, and are counted in budget statistics
• Gateway Compatibility: Accepts other Gateway provider options, unified under the providerOptions.gateway object

These capabilities make Jev easier to pass security reviews in enterprise procurement. For teams already using AI Gateway as a unified entry point, adding a decision model means no new vendor-management process is required.

Zero Data Retention (ZDR) and No Training are the two most common compliance thresholds in enterprise procurement of AI services. The former requires the service provider not to persist any user data after a request completes, usually offered as a per-request switch so enterprises can temporarily enable it when handling sensitive tickets; the latter commits that inputs and outputs won't be used to improve model weights. Both are often checked by security teams alongside audit frameworks such as SOC 2 and GDPR.

AI Gateway plays the role of a unified entry point: teams only need to maintain one set of credentials, billing, and logging systems to access multiple model vendors. By accepting the providerOptions.gateway object to remain compatible with this mechanism, Jev means organizations already using Gateway can try it without setting up new procurement or key-management processes — significantly lowering the marginal cost of adopting a new model. This is also the organizational-level precondition for a "decision model" to be adopted quickly.

What It Means for Developers in Practice

From an engineering perspective, the value of Jev 1.14 isn't how strong the model's capabilities are in themselves, but that it standardizes "AI participating in decisions." Typed output eliminates the prompt-parsing layer, while probability output provides a basis for downstream threshold judgments — you can set "route to human when confidence falls below a threshold" without having to guess at the certainty expressed by the model's words.

The two updates — image and long context — extend this decision paradigm from pure text classification to input forms closer to real business. Unstructured data such as support tickets, logistics logs, and product photos can now enter the same decision API directly.

Officials recommend that developers consult the AI Gateway documentation on decision models, Jev's practical use cases, and knowledge base resources to evaluate its applicability to their own scenarios.

One-line Summary

TypeSafe AI releases Jev 1.14, adding image and long-context input and launching on AI Gateway.

Paragraph Summary

TypeSafe AI's decision model Jev has released version 1.14 and officially landed on AI Gateway. Unlike traditional large models that generate free-form text, Jev targets structured decision tasks: the caller passes in a state and a set of questions, and the model returns typed Choice, Score, and Boolean answers with probabilities, eliminating prompt parsing and format jitter at the API level. Version 1.14 adds two key capabilities — support for images as part of the state for visual judgment, and an experimental long-context mode enabled via providerOptions.typesafe.longContext, suited to very long documents, logs, and conversation histories. On the enterprise side, Jev offers per-request Zero Data Retention and No Training, plus support for logs, reports, budget statistics, and Gateway provider options compatibility. The model name is typesafe-ai/jev-1.14.0.