Not every AI task needs a full language model response. If an application only needs to choose an option, assign a score, or decide whether something is true, a decision model can return a structured answer without generating a long explanation.
Jev, from TypeSafe AI, is a proprietary example of this approach. Its Choice, Score, and Noul functions return structured decisions with confidence scores, which makes the results easy to connect to software. Developers can ask a question, receive fixed JSON and probability values, then use those outputs to trigger an action.
The catch is that decision models do not explain their answers. They use the context provided to them, compare it with their training, and return a structured result, often a single value. That makes the context sent to the model a central part of the design.
Five open-source options for local decision-making
Several open-source models give developers a way to run this style of AI without relying on Jev. They cover different needs, from large multimodal models to compact options designed for systems with limited GPU resources.
Mammoth-1
Mammoth-1 is a 27B-parameter decision model developed by Base Labs at Baseten. It is built on Qwen3.5-27B, supports text and images, and can be self-hosted under the MIT license.
Its published Jev Decision Index 0.4 result gives it a strong position among the models compared. Mammoth-1 scored 71.24, ahead of Perplexity Decider v1.2 at 69.42 and Jev 1.14 at 62.2. Those scores come from the same evaluation, so they offer a direct comparison within that test.
vLLM Semantic Router D3
vLLM Semantic Router D3 comes in sizes from 0.6B to 27B parameters. It supports text, images, and video, and returns probability distributions that applications can use directly.
D3 fits projects that need to route requests through an AI system. An application can use it to decide which model, tool, or workflow should handle a request. D3-Mini and D3-Flash provide smaller choices when GPU resources are limited, and developers can serve the models locally through vLLM’s System One endpoint.
Perplexity pplx-decider v1.2
Perplexity pplx-decider v1.2 is a 27B decision model built on Qwen3.8. It supports classification, candidate scoring, and confidence estimates, giving developers several ways to turn model output into a software decision.
One useful feature is joint question processing, which lets developers evaluate multiple questions about the same input together. The model is released under Apache 2.0, but access to its Hugging Face repository requires authentication.
H2O-Lightning-4B
H2O-Lightning-4B takes a smaller approach with 4B parameters. It has GGUF weights for experimenting with local inference and is released under Apache 2.0.
That compact size makes it a practical option for developers testing local decision-model workflows. It also gives teams a way to explore structured outputs without starting with one of the 27B-parameter models.
Kev
Kev is one of the closest open-source replacements for Jev. It is built on Qwen models and comes in four sizes ranging from 0.8B to 27B parameters.
Kev supports yes/no decisions, multiple-choice selection, and numerical scoring with calibrated probabilities. It also provides a Jev-compatible API, allowing developers to use the TypeSafe Python SDK with a locally hosted model.
Kev is released under Apache 2.0 and allows developers to fine-tune its models on their own labeled data. That combination of API compatibility, multiple model sizes, and fine-tuning support makes it a direct fit for teams moving away from Jev.
Why structured decisions matter
Decision models skip verbose text generation and return structured outputs in JSON, along with confidence probabilities. This creates low latency and saves tokens, while the fixed format makes the result easier to extract with a library than free-form text.
That determinism is useful in applications where software needs a clear answer. Decision models can support LLM-as-a-judge systems, where chatbot answers are compared with true “golden” answers defined in advance. They can also handle support ticket triage and category mapping, since tickets often contain long text that must be assigned to the right category or team.
The tradeoff is clarity. Traditional language-model output often contains the reasoning used to reach a conclusion, but decision models make a conscious choice not to include verbose reasoning traces. There is no decision trace or log showing why the model selected its result.
That means the application must supply useful context and decide what to do with the returned value and probability. The model can provide a fast, structured answer, but the surrounding software must handle uncertainty and action without an explanation from the model.
For developers who want local control, the five options offer different paths. Mammoth-1 provides a large multimodal model with an MIT license, D3 focuses on routing across models and tools, Perplexity pplx-decider v1.2 supports joint question processing, H2O-Lightning-4B targets compact local experiments, and Kev offers the closest Jev-compatible workflow.
Based on
