TypeSafe has launched Jev, a new AI model designed for programmatic decision-making, structured AI outputs and real-time automation. Founded by OpenAI veteran and ChatGPT co-inventor Diogo Almeida, TypeSafe developed Jev over two years in stealth to address software workloads that require fast, predictable and type-safe AI decisions.

Unlike traditional large language models (LLMs) that generate text sequentially, Jev is built to make structured decisions directly inside production software. The model converts unstructured application state into type-safe probabilistic outputs through a parallel inference architecture.

Jev AI Model Replaces Traditional LLM Text Generation

Jev takes a different approach to AI inference and programmatic automation. Instead of generating a sequence of tokens, the model evaluates structured choices in parallel and returns values that conform to predefined schemas.

This design targets developers building AI-powered software, automated decision systems and production workflows where conventional LLMs can introduce latency, parsing requirements and unpredictable outputs.

TypeSafe says Jev is designed to reduce these problems by enforcing output structures at the model level. Its architecture can evaluate high-cardinality decisions with up to 255 discrete options through a two-stage scoring and selection process.

Parallel Sampling Drives Faster AI Inference

At the core of Jev is a hardware-aware parallel sampling architecture that allows multiple structured values to be evaluated simultaneously.

TypeSafe developed the model using Reinforcement Learning for Calibrated Decisions (RLCD), a training methodology focused on producing probability estimates that correspond with decision accuracy.

The company reports internal end-to-end latency of between 70 and 500 milliseconds. Published workflow evaluations also recorded execution speeds up to 193.6 times faster than selected conversational AI model baselines.

For developers, the objective is to make AI decision-making faster and easier to integrate into production code without relying on traditional text-generation pipelines.

TypeSafe Targets Low-Cost, High-Volume AI Workloads

Jev is also positioned as a lower-cost alternative for specific high-volume AI inference workloads.

TypeSafe says Jev’s input processing costs $0.042 per million tokens. Because the model produces structured states rather than conventional autoregressive text, the company says output tokens are not metered.

This pricing and architecture are aimed at applications that require large numbers of rapid AI decisions, including real-time feature extraction, automated branching logic, data processing and AI output verification.

Jev AI Model Tested on Doom Bots and Wikiracing

TypeSafe has demonstrated Jev in several real-time decision-making environments.

In one test, engineers used Jev to operate a reactive Doom AI bot across game-state structures at 10 queries per second. The company reports operating costs of approximately $7 per hour.

A separate Wikiracing AI test evaluated the model’s ability to navigate complex encyclopedia link structures. Jev selected paths through the decision tree without requiring conventional reasoning-model text generation.

These demonstrations highlight TypeSafe’s intended use cases for real-time AI agents, automated decision trees, software branching logic and large-scale data workflows.

TypeSafe Opens Early Access for Jev

TypeSafe has now opened early developer access to the Jev AI model and started onboarding engineering teams from its deployment waitlist.

Rather than competing directly with conversational AI assistants, Jev is designed around a narrower objective: bringing fast, structured and probabilistic AI decision-making directly into software applications.

The launch positions Jev as an emerging approach to programmatic AI, where models can make production decisions without generating conventional natural-language responses.