TypeSafe unveils Jev, a parallel AI for deterministic decisions

TypeSafe emerges from the shadows with Jev, a "System One" model that generates typed decisions in a parallel query rather than text. The company claims latencies in milliseconds, entry costs of $0.042 per million tokens, and production demos. The RLCD training aims for calibrated probabilities, and early access is now open to developers.

Latencies in milliseconds and entry costs of $0.042 per million tokens

TypeSafe reports end-to-end latencies ranging from 70 to 500 milliseconds from its West Coast servers, where conversational deployments fluctuate between 3 and 329 seconds. On fixed computation graphs, published evaluations announce executions up to 193.6 times faster than consensus benchmarks like GPT-6 Astra and Fable 5.1. Input processing is priced at $0.042 per million tokens, while standard conversational usage varies from $0.20 to $10. TypeSafe also indicates that it does not charge for output tokens, as parallel sampling does not perform autoregressive passes. In a load test on Doom, a bot responsive to 10 queries per second operated for about $7 per hour.

Production demos and targeted applications

Demonstrations show Jev resolving branching rules on high-speed game states and within web navigation trees. During a Wikiracing test, the course was completed in fewer steps than external models lacking reasoning, thanks to an evaluation of options that avoids generating hallucinated dead ends. According to field experiments, the targeted use cases cover real-time feature extraction, management of petabyte-sized data streams, implementation of output verification layers, and automation of branching logic when manually written rules show their limits.

Typed outputs and parallel sampling at the core of the architecture

Jev forgoes text and string generation to process an unstructured state and produce, in a parallel query, type-secured structured values. The architecture aims for deterministic code integration and is designed to prevent syntax type errors and output hallucinations. A hardware-aware parallel sampler evaluates and delivers all structured values simultaneously, under predefined output schemas that avoid parsing and safeguards often necessary with LLMs. The sampler manages up to 255 discrete options via a two-step process, with independent evaluation of options before selection. The model targets systems requiring automated deterministic logic and presents itself as a System One model for structured probabilistic decisions directly within production codebases.

RLCD training and probability calibration

The platform relies on a training method called RLCD, in contrast to the common RLHF or RLVR approaches used in conversational models or isolated programmatic verifications. TypeSafe states that it trains Jev to produce calibrated probabilities based on execution logic, so that the confidence scores returned correlate with the accuracy of the outputs. Diogo Almeida summarizes the model's interface as a boundary function call: an unstructured state as input, typed probabilistic decisions as output.

A stealth launch after two years and early access opened

After two years of stealth development led by Diogo Almeida, a veteran of OpenAI and co-inventor of ChatGPT, TypeSafe officially unveils Jev and its parallel sampling mechanism for decision automation. The company announces the opening of early access to developers and the commencement of integration with engineering teams on its deployment waitlist.