Jev is built for decisions that happen constantly inside AI applications: route a request, choose an agent action, score a condition, or flag a policy issue. It takes application state plus a predefined question and returns a typed answer with probabilities that code can use immediately.
For developers who want that workflow on local hardware or inside a self-hosted stack, there are now several open-source options worth a look. Here are nine of the best Jev alternatives available in 2026.
Last updated: September 22, 2026.
Quick comparison
| Project | Best for |
|---|---|
| Laya | Dedicated local decision models with multilingual checkpoints |
| Kev | System One compatibility plus trainable Qwen-based models |
| SemIf | Direct option scoring with Qwen or MiniCPM |
| jeff | Lightweight self-hosted System One server |
| OpenJev | DiffusionGemma decisions, image questions, and local chat |
| LocalJev | TypeScript apps with an OpenAI-compatible local server |
| OpenJev Verdict | Small browser-ready decision model with abstention |
| NanoJev | Parallel action selection for games and control loops |
| jevlike | Training a compact scorer from labeled data |
Best Open-Source Jev Alternatives
1. Laya
Laya is a local decision engine built around Jev-style choice, score, and noul questions. One forward pass handles the application state and all requested questions, returning typed answers with probabilities.
Laya ships three checkpoints. The 421M English model handles English workloads. The 322M multilingual model handles 100+ languages. A second 421M model is trained for typed-decision workloads. Laya’s Router detects the input language before inference and selects the appropriate checkpoint.
Laya can be fine-tuned on custom data. Its training package handles RLCD training, temperature fitting, evaluation, and Hugging Face publishing. Apple Silicon deployments can use Laya-MLX, and a Core ML port is published for Apple platforms.
The typed-decisions checkpoint reached 0.766 accuracy on Laya’s 2,000-decision evaluation.
2. Kev
Kev is a family of Jev-style decision models built on Qwen3.5. Each checkpoint uses a Qwen backbone, LoRA adapters, and a learned readout head to return probability distributions for choice, score, and noul questions.
Kev uses TypeSafe’s System One request format. TypeSafe Python SDK clients can connect to a local Kev server after changing the base URL.
Current checkpoints are 0.8B, 4B, and 9B. Kev-4B is the recommended starting checkpoint. Kev-0.8B uses less memory. Kev-9B reached the highest transfer-test accuracy in the current family. Kev provides the data pipeline, evaluation suites, training code, and local server for custom training.
On Kev’s current transfer test set, Kev-4B reached 0.837 accuracy and Kev-9B reached 0.852. Each released checkpoint stores temperature calibration for its probability outputs.
3. SemIf
SemIf uses existing open language models for typed decisions. It reads the scores assigned to declared answer options directly from model logits and returns typed probabilities. No answer sentence or JSON response is generated first.
Qwen3.5-4B and MiniCPM5-2B are the current reference models. SemIf runs on NVIDIA GPUs through CUDA and on Apple Silicon through MLX. Its WebGPU demo runs quantized models directly in the browser. A shared application state can be prefetched once and reused across many questions.
Qwen3.5-4B reached 0.813 balanced accuracy on SemIf’s authored decision set and 0.845 agreement on its aligned TypeSafe subset. Existing open-model deployments can use SemIf for typed option scoring with their current model.
4. jeff
jeff packages a 400M GLiFormer classifier behind TypeSafe’s System One API. Requests go to /v1/systemone, and TypeSafe SDK clients can connect through a local base URL.
The model runs on CUDA, Apple MPS, or CPU. ONNX Runtime can handle the encoder on CPU. The server provides batching, API keys, rate limits, queue controls, ONNX settings, and configurable temperature scaling.
The GLiFormer classifier handles classification-style jobs such as ticket routing, labeling, triage, and policy checks.
5. OpenJev
OpenJev runs Jev-compatible typed decisions on DiffusionGemma 26B-A4B. NVIDIA systems use vLLM. Apple Silicon uses an MLX backend. The server exposes choice, score, and noul through /v1/systemone and accepts the model aliases used by TypeSafe SDK defaults.
OpenJev can answer questions about images and exposes an OpenAI-compatible chat endpoint from the DiffusionGemma deployment. The local deployment can handle typed decisions, image inputs, and text generation from one model.
6. LocalJev
LocalJev is a Bun and TypeScript service that accepts System One requests and forwards them to an OpenAI-compatible local model server. Its default configuration uses DiffusionGemma through oMLX, and the upstream model endpoint is configurable.
LocalJev converts each typed question into a classification request, validates the returned JSON, normalizes probability vectors, and sends back a Jev-compatible response. API keys, queue limits, timeouts, chunking, and retries are part of the server.
JavaScript and TypeScript applications can keep the System One request structure. LocalJev handles communication with the local inference server, and its probability values come from the model’s structured response.
7. OpenJev Verdict
OpenJev Verdict is a 151M ModernBERT decision model for Jev-style classification. It returns the selected answer, a probability distribution, confidence, and an abstention probability for inputs that lack enough evidence.
Verdict accepts up to 24 substantive candidates plus one abstention slot. A WebGPU playground runs the model locally in the browser through ONNX Runtime Web and shows the raw probability outputs.
8. NanoJev
NanoJev is a 0.6B parallel decision model built on Qwen3-0.6B. It accepts a state, a question, and a changing candidate set, then returns the full probability distribution in one pass. Choice questions can contain 2 to 255 candidates.
The current unified-games-v1 checkpoint was trained across Maze, Snake, ViZDoom Basic, and ViZDoom Predict Position. The release contains the checkpoint, dataset, training pipeline, evaluation traces, and a persistent inference service.
On its 274-case held-out test set, NanoJev completed 4/10 Maze episodes, 8/8 Snake episodes, 128/128 ViZDoom Basic episodes, and 27/128 Predict Position episodes. Jev recorded 7/10, 8/8, 56/128, and 11/128 in the matched runs. The released checkpoint focuses on repeated action selection across these four environments.
9. jevlike
jevlike is a training kit for building a small model that chooses from a changing list of text options. Each option attends to the encoded context, and a shared scorer returns one probability for every candidate.
The default configuration learns byte embeddings from scratch. Another configuration keeps a Hugging Face encoder frozen and trains a compact scoring head on top. Training runs on CPU, Apple MPS, or CUDA.
Training starts from labeled examples. The resulting scorer can handle classification, ranking, recommendation logic, or controller decisions with a defined candidate set.
On Wikispeedia data with destination pages held out from training, the frozen Qwen2.5-0.5B configuration reached 26% top-1 accuracy. A small scratch model trained on 40,000 clicks reached 29%, and synthetic menu experiments reached about 98%.
License: MIT
FAQs
Is Jev open source?
No. Jev is a hosted TypeSafe model, and TypeSafe has not released its model weights or training recipe. The projects above provide open-source code for local Jev-style decision models and compatible servers.
Which open-source alternative is closest to Jev?
Laya and Kev are the closest matches to Jev’s dedicated decision-model setup. Kev, jeff, OpenJev, and LocalJev expose System One-compatible endpoints for existing TypeSafe integrations.
Which Jev alternatives work on Apple Silicon?
Laya has MLX and Core ML ports. SemIf includes an MLX backend. Kev runs on Apple Silicon. jeff uses Apple MPS. OpenJev has an MLX backend for DiffusionGemma.
Can these projects run fully offline?
Yes. Each project has a local deployment option. LocalJev also needs a local inference server.
Which projects work with an existing TypeSafe integration?
Kev, jeff, OpenJev, and LocalJev expose System One-compatible endpoints. Existing request and response handling can usually stay close to the current integration after the base URL changes.
Related Jev resources
- The Ultimate Jev Resource List: SDKs, Agents, MCP Servers & More
- 10 Jev Use Cases: Real Projects & Demos
