Sub-second typed classification and routing decisions with calibrated probabilities, for software branches that don't need generated text.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Calibrated Decisions at Sub-Second Latency, Without Token Generation
Technical Specifications
What the model supports
Comparison
Jev 1.13 or Jev 1.14: Pin the Version or Move Up?
Jev 1.14 is the direct successor, released October 14, 2026 with the same input pricing and the same default context limits. TypeSafe describes it as small improvements across the board, with the largest expected gain in multilingual support, but has published no numeric scores. The practical decision is whether to keep a pinned 1.13.0 whose confidence thresholds you have already tuned, or migrate to 1.14.
Choose Jev 1.13 when your thresholds, escalation rules or audit trails are tuned against it. TypeSafe's published jaggedness list also covers 1.13. Choose Jev 1.14 when you handle non-English text, need Box questions (1.14 only), or want the newest weights. Image input and long context work on both versions behind opt-in headers. Replay representative traffic before switching, since aliases such as jev-latest can move and shift scores.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Request and model routing
A single Choice question can pick the model tier for each request. LangChain's router middleware and LiteLLM's Jev Auto Router both use Jev for this. With vendor-reported latency of 70–500 ms, the routing step adds little overhead. It suits stacks that send simple requests to a small model and hard ones to a larger model. Validate routing accuracy on your own traffic.
Agent tool-call gating
Jev can check whether a tool call is risky and block it before execution, as in LangChain's AutoModeMiddleware. It can also choose an agent's next browser action in one request. Typed outputs remove the parsing step, and probabilities let you set allow, confirm and deny thresholds in code. Hard policy rules and human approval should still cover irreversible actions.
High-volume ticket triage
One support ticket can be evaluated in parallel for department, urgency and refund intent, as TypeSafe's examples show. Because questions run in isolation against a shared state, adding checks barely changes latency. This fits inbox sorting and queue assignment at volume. Because Jev is not a counting or arithmetic tool, compute order totals or dates in code first.
Moderation and guardrail screening
Noul questions return a probability that a statement is true, which suits prompt-injection checks, policy screening and comment classification before content reaches an LLM. Confidence values support escalation to review. TypeSafe's guardrails cookbook results came from jev-1.12, not 1.13, so re-derive thresholds on your own labeled examples. Keep mandatory safety rules in code.