When users ask AI to plan a trip, they may only need to say, “Help me find a suitable flight and hotel.” But to complete the task, AI must first decide: Which information needs to be checked in real time? Which tools should verify prices and cancellation or change policies? And if booking and payment are involved, at what point should the system stop and ask the user for confirmation?

As the number of models and tools grows, these decisions become increasingly important. Recently, decision models such as Jev have drawn growing attention to a broader question: Beyond answering and executing, AI also needs to know what to do next.

Today, Unisound officially launches and open-sources U2-Decision, a decision foundation model with 4B parameters in this open-source release. Positioned as the decision layer for multi-model, multi-Agent, and multi-tool collaboration, U2-Decision focuses on three questions: Who should handle the task? Can the task be executed? What path should be taken to complete it? Its three corresponding core capabilities are intelligent model routing, safety decision-making, and workflow routing.

The idea behind U2-Decision is straightforward: not every problem should be sent to the most capable model, and choosing the right model does not guarantee that the task will be completed. A truly usable AI system must continuously make trade-offs among quality, cost, speed, and risk.

Intelligent Model Routing: Choosing the Right Model for Each Task

Writing a routine email, analyzing a complex report, and answering a professional question require different levels of model capability. Even within the same category of task, the difficulty, time-sensitivity, and acceptable cost can vary. In the past, model routing often relied on preset rules: a given type of request would always invoke a fixed model. U2-Decision instead evalsuates task type, complexity, risk, model capability, cost, and latency together before selecting a more suitable model.

For example, when processing the same document, extracting dates and names may not require a complex reasoning model. But if the user asks the system to reconcile conflicting information across multiple documents and explain the basis for its judgment, stronger analytical capabilities are needed. U2-Decision is designed to recognize the difference between these tasks and route each one to an appropriate model.

Safety Decision-Making: Defining the Boundary Before Execution

AI safety issues do not arise only after an answer has been generated. Some tasks require upfront judgment about whether the request is appropriate, whether the data can be used, and whether the current model can handle the task reliably. For example, if a user asks AI to create public-facing material from a file containing sensitive information, the system cannot focus only on whether it can produce the content. It must first determine what information may be used, what should be withheld, and whether the result requires additional verification.

U2-Decision incorporates request risk, data risk, and model risk into the decision process. For high-risk or highly uncertain tasks, it can escalate the task, add verification, hand it off for human confirmation, or stop execution. In this way, safety becomes part of the task’s upfront decision-making process rather than merely a filtering step after generation.

Workflow Routing: Deciding How the Task Should Be Completed

Even after a model has been selected, a task may not be completed in a single call. In the travel-planning example above, the system needs to query real-time information, compare multiple options, verify relevant rules, and only then form a recommendation. Some steps are best suited to model-based analysis, while others require search or tool calls. When booking and payment are involved, user confirmation is also necessary.

U2-Decision extends the scope of decision-making from “which model should be called” to “which execution path should be taken.” Models, Agents, tools, retrievals augmentation, and result verification can be combined as needed. Simple tasks can be answered directly, while complex tasks can be executed step by step with verification at key points.

This is also what distinguishes U2-Decision from a conventional model selector: it is concerned not only with choosing one model from A, B, or C, but also with whether the task should proceed and where the next step should lead.

Three evalsuation Metrics Validate Decision-Making Capability

In comparative experiments against Jev and Kev-4B, U2-Decision 4B achieved the strongest results among the compared models on all three core metrics: 91.5% general-task accuracy, 99.6% safety-compliance accuracy, and 0.6 routing consistency.

These three metrics correspond to the core questions U2-Decision is designed to address in real-world use: whether it can make accurate judgments, maintain safety boundaries, and provide stable routing choices for similar tasks. The evalsuation results provide evidence in support of U2-Decision’s design.

Bringing Multi-Model Collaboration into Real-World Tasks

As an AI application connects to more models, Agents, and tools, a new challenge emerges: the capabilities are already available, but how should the system invoke them?

For users, the ideal experience is to hand a requirement to AI without having to figure out which model to choose, which tool to call, or when the result should be checked. For developers, decision-making capabilities can help systems quickly route simple tasks, allocate sufficient capability to complex tasks, and introduce necessary confirmation and verification at high-risk stages. U2-Decision is designed for this process. It does not replace the models responsible for generation and execution; instead, it continuously makes decisions as tasks begin and progress, allowing different capabilities to play their respective roles.

This open-source release is a step in Unisound’s effort to invite developers to explore applications of decision models together. We look forward to seeing more users put U2-Decision into real task workflows, test the problems it can solve, and identify where it can be further improved.

Open-source repository: http://github.com/Unisound-LLM/u2_decision_4b