Skip to content
← Back to blog
custom-modelsArticle

Local Jev 1.13 alternatives: Kev and TinyJev

Can Jev run locally? Explore Kev, TinyJev and other approaches to private decision models, with selection, fine-tuning and deployment from Syntalith.

Author

Syntalith

Published Updated 2 min read

Interest in Jev quickly produced decision-model projects that can run on your own hardware. Search results can blur the distinction between TypeSafe’s product and other developers’ tools. Kev and TinyJev are independent projects inspired by a similar question-and-answer interface.

For a business requiring local processing, this is a useful direction to test: a model reads text and selects from an agreed set of outcomes. Syntalith compares that approach with rules, classifiers and general-purpose LLMs.

What can you actually download?

ProjectWhat its authors describeWhat to verify for your deployment
TypeSafe Jev 1.13Hosted decision service, available through OpenRouterContracts, the data path through both services and processing location
KevSmall Qwen-based decision models, weights and training optionsCheckpoint, hardware, categories and adapted quality
TinyJevLocal models for choice, score and noul questions, using MLX or PyTorchVersion, device support and Polish examples
Open Alternative to JevA library reading decisions from open-weight modelsBackend differences and probability readout limits

This is a summary of author-described functionality checked on 28 September 2026. It is not a quality ranking. These projects change quickly, so a deployment should pin a specific version.

Kev: a route to task-specific training

The Kev repository describes a family based on Qwen3.5 and Qwen3.8, with variants from 0.8B to 27B. It provides models and a training process. The authors also document a local server compatible with the System One interface.

Your business still needs its own evaluation. Categories such as returns, quality complaints and missing components may overlap differently from the author’s datasets. We prepare labelled examples, separate training from evaluation and compare the model before and after adaptation.

A small adapter does not remove base-model requirements. Deployment still needs compatible weights, an engine and enough memory for actual input lengths.

TinyJev: assess the smaller footprint

TinyJev publishes 0.6B and 4B variants, with Apple Silicon and PyTorch paths. The smaller model is interesting for a personal computer or a narrow application function. Its size does not establish accuracy on Polish correspondence or speed with long documents.

We measure complete response time, memory usage and category errors. Author benchmarks do not replace evaluation in your language and under your operating rules. If a model misses the threshold, a larger variant or another approach can be tested.

Probability readout has limitations too

Open Alternative to Jev uses existing open-weight models. Its author describes different question layouts and result extraction through Hugging Face and vLLM. Depending on the backend, the available distribution may be limited, affecting how probabilities should be interpreted.

We do not describe this library as a reproduction of Jev’s weights or training. It can still be evaluated for the same business objective at an acceptable error cost. Shared data and criteria matter more than similar names.

Deployment goes beyond installation

A local endpoint needs access controls, limits, versioning and monitoring. If a decision causes a system update, the application checks authorization and agreed execution conditions. Before upgrading, rerun quality tests and retain the previous version.

We also help individuals assess these models on existing hardware. If you are planning a purchase, we start with the task and memory requirements. Syntalith selects, adapts and deploys private models, including small models for recurring decisions. Send example categories and a description of your data so we can propose a useful trial.

Match a model to the task you need it to perform

Describe where your current AI falls short. We will compare model customization options, data requirements and the cost of running the resulting system.

Private LLMs and fine-tuning
Discuss a custom model