Top 7 Open-Source Jev Alternatives for AI Decision-Making
Not every AI task needs a large language model to generate text.
Sometimes an application only needs to make a decision: Which team should handle this ticket? Is this request urgent? Should this action require human approval?
TypeSafe AI's Jev is designed for exactly these kinds of tasks. Instead of generating text token by token, it produces structured decisions such as choices, scores, yes/no judgments, and probabilities.
But Jev itself is not fully open. Its architecture, model weights, and complete training process are not publicly available.
Open-source projects are now exploring similar ideas with models that developers can inspect, modify, and run locally.
In this article, we'll look at seven open-source Jev alternatives: Laya, Nimble, Kev, SemIf, Rizzo Flow, Von, and NanoJev.
1. Laya
Laya is a purpose-built decision model based on ModernBERT, with a separate mmBERT checkpoint for multilingual use.
Instead of generating a response, Laya processes the input and directly scores possible answers.
It supports decision types such as:
Choice
Score
Yes/no judgments
Multilingual decisions
Its relatively small models, ranging from around 322M to 421M parameters, make it interesting for applications that need fast local inference.
Best suited for
Small, efficient, and multilingual decision-making.
2. Nimble
Nimble takes a different approach by building typed decision-making on top of Qwen3.5-9B.
It uses LoRA fine-tuning and reads the model's logits for the permitted answers rather than generating a complete response.
This means the system can return both the selected answer and the probability of each possible option.
One of Nimble's biggest advantages is its openness. The project provides the model, training approach, evaluation setup, and data-curation pipeline.
Best suited for
Developers interested in training and researching Jev-style models.
3. Kev
Kev is one of the closest open-source implementations to the Jev approach.
It is available in 0.8B, 4B, and 9B versions and uses Qwen with a LoRA adapter and a pointer-based decision head.
Kev can process multiple questions using the same state and return probability distributions without generating text.
It also implements the System One API, making it easier to use as a local alternative in applications designed around the Jev interface.
Best suited for
Jev-like local deployments and developers looking for different model sizes.
4. SemIf
SemIf takes perhaps the simplest approach.
Instead of training a dedicated decision model, it uses existing open models such as Qwen and reads the logits assigned to predefined answers.
For example, instead of asking an LLM to explain whether a support request is urgent, SemIf can directly calculate the probability of:
Urgent → 82%Not urgent → 18%
This avoids generating unnecessary text.
Best suited for
Experimenting with decision models without training a new model.
5. Rizzo Flow
Rizzo Flow focuses on local-first inference and hardware flexibility.
It uses open models with llama.cpp-based serving and converts possible answers into constrained choices. The system then reads the probability of those choices directly.
One useful feature is its ability to process a state once and reuse it across multiple questions.
It also provides a System One-compatible HTTP interface.
Best suited for
Local AI deployments and flexible hardware environments.
6. Von
Von is a compact System One-style model built around a roughly 395M-parameter ModernBERT architecture.
It uses candidate representations to produce probabilities for different decisions instead of generating text.
Von supports:
Choice decisions
Yes/no decisions
Score decisions
Multiple questions in one forward pass
Its small size makes local inference particularly interesting.
Best suited for
Lightweight decision-making on local hardware.
7. NanoJev
NanoJev is a lightweight Jev-inspired model built on Qwen3-0.6B.
It uses dedicated decision heads for different types of structured decisions, including choices, Boolean judgments, and scores.
The model can process multiple states and questions together, making it useful for experimenting with small, trainable decision models.
Best suited for
Developers looking for a small and trainable Jev-style model.
Comparing the 7 Jev Alternatives
Project | Main Approach | Approx. Size | Best For |
Laya | ModernBERT / mmBERT | 322M–421M | Multilingual decisions |
Nimble | Qwen + LoRA | 9B | Training & research |
Kev | Qwen + pointer head | 0.8B–9B | Jev-like local deployment |
SemIf | Direct model logits | Model-dependent | Easy experimentation |
Rizzo Flow | Local constrained scoring | 1.7B–4B | Local serving |
Von | ModernBERT + decision head | ~395M | Lightweight inference |
NanoJev | Qwen3 + decision heads | 0.6B | Small trainable models |
The projects differ mainly in how they produce decisions.
Laya and Von use compact encoder-style architectures. Kev, Nimble, and NanoJev adapt Qwen models with specialized decision mechanisms, while SemIf and Rizzo Flow focus on extracting probabilities directly from existing models.
Which Jev Alternative Should You Choose?
The right option depends on what you need.
For multilingual workloads: Laya
For a close Jev-style implementation: Kev
For lightweight local inference: Von or NanoJev
For training and research: Nimble
For experimenting without specialized training: SemIf
For flexible local serving: Rizzo Flow
The important difference is that these projects let developers control the model and inference stack locally, rather than relying entirely on a proprietary decision API.
Conclusion
Jev highlights an important idea in AI: not every application needs text generation.
For many agent workflows, the required output is simply a decision, probability, or classification.
Open-source projects such as Laya, Kev, Von, Nimble, SemIf, Rizzo Flow, and NanoJev are exploring different ways to make this approach accessible locally.
As AI agents become more complex, small decision models could become an important layer between raw data and large language models.
Comments