Kai 1: Decisions, Not Completions

Most steps in production software don't need a paragraph. They need an answer the software can act on: which tool to call, whether the evidence is enough, how severe a risk is, whether a human must sign off.

Today we're introducing Kai 1, Hanzo's decision model. Give Kai the facts of a situation and a question with listed options. It returns a typed answer (a choice, a score, a yes/no, or a calibrated "not sure, defer") with a probability for every option.

Kai 1 is live on the Hanzo API. Its weights are closed: Kai is served by API only.

Kai reads eight kinds of evidence about a vehicle and returns ready, needs review or not ready, each with a probability, or defers to a human. — Figure 1 · One shared state, one listed question, one typed answer with a probability for every option. Probabilities are illustrative.

Language models reason. Kai decides.

When you ask a language model to make a call, you get text back. You have to parse it, trust a confidence number the model made up, and hope the output stays in bounds.

Kai is built for the steps where that isn't good enough. Its answers are bounded, its probabilities are calibrated, and there's nothing to parse. Every Kai answer is one of four types:

Choice: one option from a listed set, with a probability for each
Score: a rating against levels you define
Predicate: a yes/no with its probability
Defer: Kai isn't confident enough and hands the call to a person or a rule

Evidence stays attached

A decision shouldn't care whether its evidence started as text, an image or a sensor reading. Kai is designed to take mixed evidence and keep its origin attached to every answer.

A number without provenance is just a number.

Fifteen kinds of evidence (text, code, image, video, audio, radar, lidar, thermal, sensors, telemetry, time series, graph, SysML, simulation and documents) pass through modality encoders into one shared evidence state, which feeds Kai. — Figure 2 · Text, images, sensor readings and more are encoded into one shared evidence state that Kai reads.

Many decisions, one pass

Agent loops ask one question at a time. Kai is designed to run a whole decision program at once: one shared state, many decisions, with dependencies spelled out. Confident answers are locked in first, and only the uncertain ones get more work.

The goal is that a program with thousands of decisions shouldn't need thousands of serial model calls.

A grid of 100 decision nodes that all start masked. 87 resolve on the first pass, 11 on a second pass, and a constraint pass resolves the last 2 conflicts. — Figure 3 · Confident decisions settle on the first pass; only the uncertain ones are refined. Counts are illustrative.

Every decision can be replayed

Each Kai decision records the program version, evidence fingerprints, model revision, full probability distribution, approvals and execution trace. When someone asks why the system did something, you can show them.

Ten kinds of evidence about vehicle 17, from camera inspection and CAN telemetry to SysML architecture and requirements, flow into a decision program, then Kai, then policy and solver checks, then human approval, ending in a decision package. — Figure 4 · From evidence to a decision package: the record that lets a decision be replayed. Illustrative example.

Where Kai fits

Agents: model and tool selection, when to continue or stop, when to ask a human
Business workflows: lead scoring, intent detection, next best action, escalation
Physical systems: vehicles, robotics, manufacturing, energy and infrastructure, where decisions run on sensor data

Part of the Hanzo stack

Enso routes each task: writing goes to Zen6, decisions go to Kai 1. — Figure 5 · The Hanzo stack. Zen writes, Kai decides, and Enso picks which one to call.

Generation and decision are different jobs. Zen6 is our family of open-weight generation models. Kai 1 makes bounded, calibrated decisions. Enso routes each task to the one it needs.

Where it runs

Kai runs on the Hanzo API, at POST https://api.hanzo.ai/v1/decisions, for $0.021 per million input tokens with output free. Its weights are not published, and it is not offered for self-hosting.

What's next

We're working on seven open research questions, and we'll publish results only once they can be reproduced:

Decisions over extremely large option sets
Multilingual decisions
Shared-state decoding
Decision diffusion (refining only uncertain answers)
Multimodal evidence
Calibration and knowing when to abstain
Consistency across connected decisions

Kai 1's first results are published. On the 12 headline suites of a frozen decision harness, Kai 1 averages 85.7% accuracy against Jev's 78.0%. It is more accurate on 11 of 12 suites, behind on jailbreak (91.75% against 94.0%), and its calibration error is lower on 8 of 12. Those runs were measured on CUDA; production answers on CPU. Kai 1 is not yet competitive on dependent multi-question programs, on negated yes/no questions, at 150 or more options, or on calibration for typed decisions and jailbreak, and at 1,000 or more options its accuracy is near zero. Every table, with the run each number comes from: hanzo.ai/kai and the Kai docs.

Kai 1 · Hanzo API

Stop asking one model to do everything

Call Kai through the Hanzo API today.

Explore Kai →Start building with Hanzo Meet the Author ↗

Kai 1: Decisions, Not Completions

Antje Worring