jev model card

What This Model Card Can Establish

This is an operational snapshot assembled from TypeSafe’s public model, interface and jaggedness documentation. It is not a vendor-issued safety report, an independent benchmark, or a disclosure of Jev’s weights and training data. Mutable facts carry a verification date because aliases, limits, price and provider availability can change without this page changing automatically.

The model’s job is stable across those mutable details: consume text state and typed questions, then return bounded decisions and probability distributions. For the full application boundary, read how Jev works; for a task-selection guide, read when not to use Jev.

Model Identifiers and Release Channels

An alias is a deployment pointer, not a model version. Using jev-latest in a request does not eliminate the need to record the resolved model returned by the service. If a threshold or policy was fitted on one version, silent alias movement can change coverage and error rates while application code remains unchanged.

Input, Context and Throughput Limits

TypeSafe documents text input and a 64k-token request budget across state and questions, with a separate 32k limit for state plus the longest individual question. These are acceptance limits, not a recommendation to fill the window. The Jev 1.13 jaggedness page says irrelevant long state can reduce accuracy. A production state projector should select the smallest evidence that preserves the criterion and measure quality as state grows.

The model page lists 250,000 tokens per second and 1,200 requests per minute for the direct service, while explicitly saying rate limits are adjusting dynamically. Capacity planning must therefore read current response headers or account documentation, handle 429s, bound retries and reserve room for bursts. OpenRouter or Cloudflare quotas are provider-specific and should not be copied from TypeSafe’s direct limits.

Price the Serialized Workload, Not the Visible Sentence

At the documented direct rate, one million input tokens costs $0.042. One million calls averaging 500 input tokens consume 500 million tokens, or $21 in model input charges. That arithmetic excludes retries, provider markup, orchestration, observability, human review, fallback models and engineering work. The pricing guide develops those scenarios and compares provider boundaries.

Questions consume input as well as state. Independent questions can share one state in a request, which may avoid repeating the state, but batching changes request size and failure scope. Measure the actual serialized request and end-to-end service rather than multiplying a guessed prompt length by the model price.

Known Jev 1.13 Behavior Belongs in the Model Card

These are not edge notes to hide below a benchmark. They determine whether an application is correctly designed. The complete limitations guide connects each weakness to a regression fixture and mitigation.

Language and Data-Use Claims Need Deployment-Specific Checks

TypeSafe says English is Jev’s strongest language and that requests and responses are not used to train the model. The first claim requires per-language evaluation before rollout; the second answers training use but does not by itself specify retention duration, residency, subprocessors or a customer’s regulatory obligations. Review current contractual and provider documents for those questions.

If a gateway serves Jev, assess that provider’s data path separately. Model identity does not make TypeSafe’s direct-service terms automatically apply to Cloudflare, OpenRouter or another intermediary. The platform guide keeps access method and model behavior as separate columns.

A Model-Card-Based Release Gate

1. Pin the candidate model and record provider, SDK and endpoint versions.
2. Replay a held-out set containing common, boundary, minority-slice and adversarial examples.
3. Compare class errors, calibration, coverage, latency percentiles, rate limits and cost with the current version.
4. Shadow or dual-run on current traffic without changing actions.
5. Approve thresholds separately from approving the model; record rollback criteria.
6. Monitor resolved model IDs and rerun the gate when any alias moves.