TypeSafe AI Launches Jev, a Fast Non-Text AI Model for Automated Software Workflows

TypeSafe AI, an artificial intelligence startup emerging from two years in stealth, has released its initial model, Jev, into early access. Designed as a specialized System One Model, Jev focuses on producing structured, machine-readable decisions rather than generating conversational text, aiming to streamline automated software workflows.

The product launch was first reported on Hacker News, following development led by a former OpenAI researcher who previously contributed to instruction-following methods behind ChatGPT. According to the company, standard large language models present major latency and reliability bottlenecks when integrated into software code, largely due to open-ended string generation, type formatting errors, and unpredictable hallucinations.

To eliminate those limitations, TypeSafe AI built a hardware-optimized technology architecture equipped with a parallel sampler and a novel training technique called Reinforcement Learning for Calibrated Decisions (RLCD). Instead of generating outputs token-by-token sequentially like conventional chatbots, Jev processes all outputs in a single query. The approach completely discards text string generation in favor of pre-defined, type-safe structured data accompanied by calibrated confidence probabilities.

The company reported that Jev achieves end-to-end response times between 70 milliseconds and 500 milliseconds, representing a 40x to 200x speed increase over standard frontier models running on System One tasks, which typically exhibit response times ranging from 3 to 329 seconds. In multi-step programmatic workflow evaluations, TypeSafe AI stated that Jev ran up to 193.6 times faster and reduced operational costs by up to 444.6 times compared to conventional model wrappers.

By forfeiting open-ended string outputs, the model guarantees schema alignment without type errors, making traditional structural hallucinations mathematically impossible. Every result provided by Jev includes explicit confidence metrics, allowing host applications to parse calibrated probabilities before executing automated conditional logic or branching decisions.

TypeSafe AI validated Jev using a newly created evaluation harness designed to test AI performance within programmatic compute graphs. The testing framework compared Jev’s decision accuracy against ensemble reference outputs from frontier systems, including GPT-6 Astra, GPT-5.6 Terra, and Fable 5.1. The benchmark results positioned Jev at the leading edge of performance efficiency for tasks such as programmatic routing, data classification, and verification guardrails.

To demonstrate real-time capabilities, the startup showcased Jev powering an automated bot executing instructions in Doom using raw state data structures at ten queries per second, as well as a traversal agent navigating complex Wikipedia links with choice sets up to 255 items. TypeSafe AI is currently bringing developers off its waitlist for early access, targeting enterprise use cases in data processing, automated verification, and real-time application logic.