jev-preview

Overview. Jev is a proprietary text-input model from TypeSafe AI, trained with RLCD to return calibrated decisions. Each request returns typed answers keyed by question name, such as a choice with probabilities, a score, a noul probability, or box coordinates, rather than generated text. The typesafe/jev-preview slug is an alias that currently points to jev-1.14.0, the same release as jev-latest.

Key strengths. The vendor's model documentation reviewed for this model contains no benchmark results, so this section makes no quality claim. The vendor documents a structured interface: choice and score answers include probabilities, and the default per-request context window is 64k tokens. English is the primary training language, where the vendor reports accuracy is best.

When to use it. Use Jev when your application needs a structured decision over text, such as selecting one option from a fixed set of choices or returning a score against a defined legend. It suits pipelines that parse answers programmatically and do not need prose. Avoid it for chat, drafted replies, code generation, or any task where a written explanation is the expected output.

Limitations. Jev is not a chat or completions model, and it does not write replies, produce code, or explain its reasoning. The default context is 64k tokens per request and 32k per question, and a 262k window is available only when the vendor's long-context header is sent. Parameter count, architecture, training cutoff, and output limits are not published, and because the alias moves with new releases, pin jev-1.14.0 when results must be reproducible.

Sovereignty note. TypeSafe AI is recorded with a United States country of record, and its creator record flags CLOUD Act exposure, so this model should not be presented as Sovereign. The vendor states that customer requests and responses are not used for training, and that the same weights serve every account. Data residency for the endpoint has not been verified, so no residency claim is made.