NEAR AI and Levanto Labs Bring Verifiably Private Decision Models to Life

Most people think about AI in terms of answering questions, summarizing documents, writing code, or creating content, but decision making is becoming more important as agent adoption continues to scale.

Some of these decisions will save companies millions of dollars. Which model should handle this prompt? Is this request simple enough for a smaller open-source model, or does it require a frontier model? Should an agent be allowed to take an action automatically, or should it ask for approval first? Does a request violate a company policy, trigger a compliance requirement, or need to be escalated to a human?

That is why we are excited to partner with Levanto Labs to bring Sage, its proprietary decision model, to NEAR AI's confidential computing infrastructure and explore what's possible when the decision layer can remain private.

Levanto describes Sage as the first dedicated decision model, built specifically for machine-to-machine decision making rather than long-form generation. Sage can classify inputs, make structured decisions, return calibrated confidence scores, and also supports images, grounding, and reasoning.

The early benchmark results are impressive, with Sage previously ranked #1 on JevBench, ahead of models from Microsoft, OpenAI, and Jev. For anyone interested in digging deeper into how Sage performs across decision-model benchmarks, the full results are available on JevBench.

Use case: prompt classifier

One of the use cases we are most excited about internally is using Sage as a verifiably private prompt classifier and model router across NEAR AI.

Today, many companies send nearly every employee request to the same expensive frontier model, even though a huge percentage of those prompts could be handled just as well by much cheaper open-source models. A basic extraction task, a simple classification request, and a difficult coding problem don't need the same model, but employees shouldn't have to navigate an increasingly complicated model landscape every time they ask AI for help.

A smart router can automatically choose the best model for each request based on capability, cost, latency, privacy requirements, and confidence. At enterprise scale, where companies may process millions of prompts, routing more requests to lower-cost open-source models could save millions of dollars while still reserving frontier models for the tasks that actually require them.

But the problem is the router has to read the prompt before it can decide where to send it. This means the host of the decision model can access the plaintext request, so the prompt is no longer private, even if the model it eventually reaches is. You can't really have private inference if the system making the routing decision sees the data first.

By running Sage on NEAR AI's confidential infrastructure, we can keep that classification step private as well, which gives enterprises a path to combine the economics of open-source models with the privacy guarantees required for sensitive workloads.

Beyond routing

The same idea extends well beyond routing, like allowing private decision models for programmatic compliance, transaction screening, agent permissions, and other situations where software needs to evaluate sensitive information before deciding whether an action should happen.

This is what makes the partnership with Levanto exciting to us. We are not simply adding another model to an inference platform but instead exploring a private decision layer for AI systems that helps companies decide which model to use, which actions agents should be allowed to take, and when humans should stay in the loop, while keeping the underlying data protected.

Over time, we also plan to explore broader distribution of Sage through NEAR AI and third-party channels, making this new class of decision model accessible to more developers building agentic systems.

Learn more about Levanto at levanto.ai.