Kitsuno runs open Laya model in shadow next to closed Jev API

On 3 October, Kitsuno published an engineering post detailing its decision to run an open-source model alongside its current closed infrastructure. The company is evaluating Laya, an open-weight model from Nandakishor Mukkunnoth and ConvAI Innovations, against its existing closed tool, Jev. This dual setup is a deliberate experiment in architectural sovereignty, not a simple upgrade path. The team wants to see if an open model can handle the specific nuances of European labor market data without leaking sensitive information to external servers. The stakes are high because the current system works, but it sits in a jurisdiction that complicates compliance. By running both systems in parallel, the engineers can measure the gap in real time without disrupting the live service that handles job matching requests every day.

The move is driven by a strict internal principle: data must stay in European data centers.

Jev runs in the United States and is closed source. Laya is available under the Apache 2.0 license and can be hosted locally. This aligns with Kitsuno's ninth principle, which mandates that data remains within Europe. The legal and operational implications of this distinction are substantial for a company dealing with personal data in the EU. Sending queries to a US-based API creates a dependency that the leadership wants to eliminate. Local hosting ensures that the model weights and the data they process never cross borders, a requirement that is becoming increasingly common in regulated industries. It is a fundamental shift from renting intelligence to owning the infrastructure that generates it.

The performance gap is significant but manageable. In a small staging test, the untrained Laya correctly identified the right country for 26 percent of job ads. It got the right work mode for 50 percent. By comparison, Jev achieved 89 percent accuracy on the same metrics. On a broader test set, Laya answered 44 percent of questions correctly, while Jev answered 94 percent.

The Shadow Mode Approach

Kitsuno is not replacing Jev immediately. Instead, it has deployed Laya in a "shadow" configuration. The open model processes the same real-world requests as Jev, but its outputs are not used for final decisions. This allows the team to measure performance in a live environment without risking production stability. The goal is to train Laya specifically on Kitsuno's unique data until it matches or exceeds the closed model's performance.

A key constraint in this process is the prohibition on using Jev's outputs as training labels. TypeSafe's customer agreement, updated on 23 September, explicitly forbids using Jev's output to perform model distillation or to train a model to imitate the service. Kitsuno interpreted this strictly. The database role for the Laya worker cannot even read Jev's answers. This means the team had to build a new labeling pipeline from scratch, which slowed the work but ensured compliance.

The current Jev stack is efficient in terms of cost. Since 23 September, it has handled 35,812 logged calls using 76 million input tokens for a total cost of USD 3.68.

The primary motivation for switching is not financial savings at this volume, but rather architectural sovereignty and the ability to customize the model for specific business needs. The cost per call is negligible, so the business case rests entirely on control. If the team can make the open model accurate enough, they gain the freedom to modify the logic without waiting for a vendor update. This independence is the core value proposition of the project. It transforms the AI component from a service dependency into an internal asset that can be iterated on by the engineering team directly.

Context from the Industry

This trend of replacing closed APIs with open weights is gaining traction. Aleph Alpha released Kolibri on 3 October, a sovereign open-weight model with 78B total parameters and 3B active parameters. It is designed for regulated industries and supports context lengths of up to 1M tokens. Like Laya, Kolibri is available under the Apache 2.0 license, allowing companies to run it on-premise without sending internal data to third parties.

Meta AI Research also highlighted the utility of open models for complex tasks. The team partnered with mathematicians to solve open research problems using Muse Spark 1.1 and 1.2. The papers clearly marked which passages were drafted by AI and which by researchers, demonstrating a transparent workflow that open models facilitate.

For many developers, the choice between closed and open models is no longer just about accuracy. It is about control, compliance, and the ability to iterate. As Kitsuno notes, the training is the product. An open model is a starting point that can be taught specific questions, whereas a closed API is a black box that cannot be modified for your specific use case.

The field is moving toward a hybrid approach where deterministic checks and agentic AI work together. The open-source community is providing the tools to build these stacks locally, reducing dependency on external providers and ensuring that sensitive data never leaves the organization's infrastructure.