Decision model comparison · Sources checked September 27, 2026
Jev vs AnyJev
Both return structured decisions and probabilities. Jev is a ready-to-use hosted model; AnyJev is an open-source toolkit that turns an open LLM you run into a decision system.
Hosted model · TypeSafe
Jev
Send a state and typed questions to an API. Jev returns answers and probabilities without setting up a model or fitting a head for each question.
Open-source toolkit · Nokia Applied Research
AnyJev
Use an open LLM for choice, yes/no, and score decisions. Start without labels at L0, then fit calibration or a question-specific head as labels arrive.
At a glance
The practical differences
They answer similar kinds of questions, but the model you operate and the work needed to trust its probabilities differ.
| Aspect | Jev | AnyJev |
|---|---|---|
| Delivery | Hosted production API | Python toolkit over your own open LLM; supports transformers and vLLM serving |
| Outputs | Choice, Noul (yes/no), and Score; multiple questions in one request | Choice, yes/no, and score decisions with probability distributions |
| Day one | Use the API without task-specific labels | L0 runs without labels; it corrects option-position and label-prior bias |
| Confidence | Calibrated probabilities for downstream decisions | L0 is not calibrated; L1 fits temperature with labels, while L2 fits a readout head |
| Task labels | No per-question head fitting required | L1: about 100–500 labels; L2: about 100–300 labels per question and model |
| Operations | API integration; model infrastructure is hosted | You operate the base model and serving stack; L2 needs hidden states |
| Availability | Commercial hosted service | Apache-2.0 project code; base-model licenses still apply |
AnyJev method
What its three levels add
AnyJev can start with no training data, but each level offers a different probability contract.
L0
No labelsAverages over option order and corrects label priors. It improves raw-logit decisions but does not calibrate uncertainty.
L1
100–500 labelsAdds temperature scaling to L0 for question-specific probability calibration. It does not change answer ranking.
L2
100–300 labelsFits a closed-form head on an intermediate hidden state, without changing the base model's weights. The head belongs to one question and model.
The README also describes a raw-logit mode. Its letter-based readout currently supports at most 26 options; span readout is on the roadmap.
Published evidence
Read the results in their own context
AnyJev reports better agreement with teacher labels after fitting an L2 head on Qwen3-8B. This is useful evidence for the method, not a measured Jev-versus-AnyJev production result.
The repository quotes a Jev result from its authors on the same dataset, but says it did not rerun Jev. JevBench scores shown elsewhere on this site use another benchmark. Neither set should be combined into a single head-to-head score.
AnyJev-reported · Qwen3-8B
64.7%
L0 accuracy
77.1%
L2 accuracy
LocalLLaMA/typed-decisions: 20 questions, 300 labels per question, 2,000 held-out decisions. Accuracy means agreement with a teacher LLM, not independent ground truth.
Choosing a path
Which one fits your workflow?
Choose Jev for a managed API
Use it when you need typed decisions and probabilities now, want to ask several questions in one call, and prefer not to serve an LLM or collect labels for each question.
Explore the Jev modelChoose AnyJev for model control
Use it when you can host an open LLM, need control over the base model and deployment, and can validate or fit probabilities on your own questions.
Read the AnyJev levelsJev vs AnyJev FAQ
Is AnyJev the same project as Jev?+
No. AnyJev is an independent open-source research project from Nokia Applied Research. It is not affiliated with TypeSafe or this site.
Can AnyJev work without training labels?+
Yes. L0 needs no labels, but its probabilities are not calibrated. L1 and L2 need labeled examples for each question to improve calibration or fit a head.
Is AnyJev more accurate than Jev?+
There is no independently rerun, like-for-like production comparison in the cited sources. AnyJev publishes results for its Qwen models and quotes a Jev result reported by Jev's authors. Test both on your own decision set before choosing by accuracy.
Source material checked September 27, 2026. AnyJev and TypeSafe belong to their respective owners. Published figures may change with models, datasets, and serving setups.
