Jev AI Hub
Start Learning

Learn

Jev 1.14 and SDK Migration Guide

Keep a versioned model for established text workflows. Upgrade SDKs deliberately, then evaluate experimental images and Box separately.

Quick answer

Keep a versioned model for established text workflows. Upgrade SDKs deliberately, then evaluate experimental images and Box separately.

Choose the model separately from the SDK

As checked on October 20, jev-latest selects jev-1.13.0 and jev-preview selects jev-1.14.0. An SDK upgrade does not change the stable alias. Pin the model ID when comparing results, and log the returned model.

CapabilityJev 1.13Jev 1.14
Choice, Score, Noul text questionsAvailableAvailable
Images with the experimental headerAvailableAvailable
Box localizationNot supportedExperimental
Default context64k request / 32k per question64k request / 32k per question
Long context with its separate opt-in header262k / 262k262k / 262k

The published input rate remains $0.042 per million tokens for both versions, with free output. Estimate image workloads from returned usage rather than assuming a fixed text-token equivalent. See pricing.

JavaScript and TypeScript

Install @typesafe-ai/sdk@0.7.0 on Node.js 22 or newer. Review imports of request payload types: this release stops exporting them. Keep Score criteria as ordered arrays. On Node.js 22.15+, bodies of at least 1 KiB use zstd; configure compression: false for endpoints that cannot decode it. Usage token counts can be null: missing measurements are not zero cost.

Follow the TypeScript tutorial for a text request before adding image assets.

Python

Pin typesafe-sdk==0.7.5 for the examples reviewed here. It adds Image assets and Box, and enables zstd request compression by default. A compatible local server or gateway may not support the new transport. Test a minimal request before moving traffic. Versions 0.7.3 and 0.7.4 also include TLS concurrency and serialization fixes.

See the Python tutorial. An older pinned SDK can remain useful while testing an upgrade; it should not be described as the latest release.

Validate before switching traffic

  1. Save a labeled set from your workload, including ambiguous and absent-target cases.
  2. Run the old and candidate models with identical questions.
  3. Record model, answers, confidence, usage, latency, and errors.
  4. Check false positives and review rates. Recalibrate thresholds.
  5. Keep the previously tested model and dependency lockfile available for rollback.

No live API comparison was performed for this guide.

Try the experimental capabilities

Images covers uploading a photo. Box explains localization and absent-target handling. Both require TypeSafe-Danger-Zone-Images: true. The long-context header is a separate switch.

Sources

  1. ModelsTypeSafe · accessed 2026-10-20 · documentation
  2. ImagesTypeSafe · accessed 2026-10-20 · documentation
  3. Python changelogTypeSafe · accessed 2026-10-20 · documentation
  4. JavaScript changelogTypeSafe · accessed 2026-10-20 · documentation