TypeSafe AI launches Jev, claiming up to 200x faster inference without hallucinations

TypeSafe AI's approach departs from the architecture that has defined large language models. Conventional LLMs generate text token by token — a sequential sampling method where errors can accumulate and inference time scales with output length. Jev, by contrast, has its output structure fixed in advance and computes all outputs in parallel in a single query. The company says the result is 70〜500ミリ秒 for inference, compared with 3〜329秒 for conventional LLMs at equivalent quality.

The startup also introduced a training method it calls RLCD (Reinforcement Learning for Calibrated Decisions), which sits alongside RLHF and RLVR. Where those optimize for human-preferred dialogue or correct reasoning, RLCD optimizes for calibrated decisions. Jev outputs both a confidence score and uncertainty for every response, which could change how downstream systems decide whether to trust an AI's answer.

The stakes hinge on whether TypeSafe AI's benchmark claims hold up outside its own testing. The company used GPT-6 Astra and Fable 5.1 averages as the reference point and applied an OSS adapter to the comparison LLMs, but those results are self-reported and the model is only in early access. Developers working on automation — particularly those who have been blocked by type errors or hallucinated tool calls — are the audience the company is courting.