Laya AI Model: The Free, Open-Source Jev Alternative

The Laya AI model is an open-weight decision engine from Convai Innovations. It came out on September 18, 2026 under the Apache 2.0 license. Like TypeSafe’s Jev, it never writes a sentence. Instead, it returns typed answers with probabilities attached. The difference is price: you can download it and run it yourself for nothing.

Laya is what its makers call a System 1 decision model. You hand it a state, such as an email, a support ticket or a JSON record. Then you ask it typed questions, and it answers each one in a single forward pass.

For instance, there are three answer types. A choice picks one option from a list, and a score ranks something on a scale. Finally, a noul question returns a yes or no probability. Because nothing is generated, there is no free text to parse and no JSON that can break halfway through.

In fact, Laya is not a chatbot at all under the hood. The English checkpoint pairs a ModernBERT-large encoder with a small two-layer decision head, for 421 million parameters in total. In other words, it is tiny next to the language models it competes with on these tasks.

Overall, Convai ships three checkpoints. The English root model reads 512 tokens, while a 322-million-parameter multilingual version built on mmBERT reads 1,024 tokens across more than 100 languages. A third checkpoint is fine-tuned for typed decisions. Meanwhile, training uses a method called RLCD, which rewards the model only for reporting honest probabilities.

Above all, the search interest this week sits on one comparison. According to AI Weekly, Laya answers a single question in 32.8 milliseconds on an Nvidia T4. Jev 1.13.0 needs 236 to 276 milliseconds for the same job. That works out to roughly 7.8 times faster.

Laya Against Jev, Metric by Metric

Here is how the published figures line up.

In short, the table flatters Laya in some rows and not in others. Still, it is the first open alternative with numbers anyone can reproduce on their own hardware.

First, the headline accuracy needs an asterisk. As AI Weekly points out, the 0.766 score belongs to a checkpoint fine-tuned on that benchmark’s own training split. Out of the box, the base models score about 0.36, barely above a random baseline of 0.318.

Second, wide option lists are another weak spot. On Banking77, with 77 possible labels, Laya scores 0.425 while Jev reaches 0.870. In addition, the raw model is overconfident until you fit a temperature per question, and its short context window means long documents need chunking. Flowtivity also tried it on a four-core CPU server and saw a median of 49.4 seconds per prediction, so a GPU is effectively required.

For developers, getting started is also simple. The weights live on Hugging Face under convaiinnovations/laya, the code sits on GitHub, and a pip install pulls the package. There is also a free demo Space and a Kaggle notebook that fine-tunes the model in about four hours on two T4 cards.

Therefore, that openness is the whole argument against Jev. TypeSafe still admits developers through a waitlist, whereas Laya is already public. For a closer look at the paid rival, see our report on the Jev AI model and its $40 million launch.

Searches for Laya local and Laya MLX jumped this week, and there is a reason. An independent developer ported the model to Apple’s MLX framework. The laya-mlx project reports 13.4 milliseconds for a short English decision on an M3 Max. The multilingual checkpoint gets that down to 7.4 milliseconds.

Note that the port is not an official Convai release. Even so, it runs without PyTorch or any cloud API. If you already keep models on your own machine, our guide on how to run an LLM locally covers the hardware side.

Convai Innovations is led by founder and CEO Nandakishor Mukkunnoth. The company published early research on the idea in March 2025, followed by a paper formalizing RLCD that September. Laya is the first product built on that work.

The pitch is aimed at narrow, high-volume jobs. For example, Convai lists invoice processing, ticket routing, moderation and email triage. As a result, the realistic buyer is a team that can label its own data and owns a GPU, not someone hoping for zero-shot magic.

Plugging a decision model into an agent? Start with our list of the top agentic AI frameworks and pick the one that fits your stack. A fast decision model slots in best at the routing and tool-selection steps.

Sources

AI Weekly, Hugging Face, Flowtivity, eesel AI, Convai Innovations, laya-mlx on GitHub.*Photos: Homepage of Hugging Face website magnified on logo by Jernej Furman, CC BY 2.0, cropped and resized; M3 MacBook Pro 14 inch Space Grey model by Kyu3a, CC BY-SA 4.0, cropped and resized.*