Kai 1 · decisions, not completions.

Kai

Decision intelligence.Not another language model.

Kai takes explicit state and a listed question and returns a choice. The coding agent, Hanzo Bot, and Hanzo Team call it for the model, the context, the tools, and the stop.

Explore the research View benchmarks

kai · multimodal · typed · calibrated · parallel · self-hostable

vehicle.readiness@3example · scripted

evidence

Example: a scripted scenario, not model output. Switch a source off and only the decisions that read it run again.

Bounded answers

Language models reason. Kai decides.

Generative models are good at open-ended work. Most decisions inside a real system are not open-ended.

Which tool should run?
Which component failed?
Is the evidence sufficient?
How severe is the risk?
Which option satisfies the constraints?
Should execution continue?
Does a human need to approve this?

Each has a bounded answer space. Kai is built for them.

hanzo decision

zenopen-ended reasoning and generationkaibounded judgment and decision stateembedretrievaltsforecastingvisionperceptionaudiospeech and acoustic perceptiongraphstructural evidencesolvermathematical correctnesspolicyauthority

 
hanzo decision composes all of them

Evidence

Everything is evidence.

Text. Code. Images. Video. Audio. Sensors. Telemetry. Time series. Graphs. SysML. Simulation. One decision plane over all of it.

evidence plane

text
code
image
video
audio
radar
lidar
thermal
sensors
telemetry
time series
graph
sysml
simulation
documents

shared evidencekai

A decision should not care whether its evidence started as prose, a thermal image or a CAN frame. Specialized encoders produce the evidence, and Kai reasons over one shared representation of it.

New sensor tomorrow? Add an adapter. The Decision Program does not change.

Sensors

Sensors are first-class.

Kai does not treat telemetry as text. A measurement carries its meaning with it.

CAN / J1939
IMU
accelerometer
gyroscope
temperature
pressure
voltage
current
vibration
GNSS
radar
LiDAR
thermal
RF
acoustics
network telemetry
industrial buses
custom sensors

A number without provenance is just a number. Kai keeps the evidence attached.

measurementexample

value104.2unit°Cchannelcoolant.temperaturesample_rate100 Hztimestamp11:42:13.428uncertainty±0.4 °Ccalibrationvalidsourcevehicle-17provenancesigned stream

Change

Decide on what changed.

A sensor stream can hold millions of samples. A decision should not have to reread all of them.

Specialized temporal models turn raw streams into evidence. Kai decides what that evidence means for the system.

coolant.temperature · summaryexample

current104 °Ctrendrisingderivative+1.8 °C/minbaseline+4.2σforecastlimit crossed in 6 minqualityvalid

Parallel

Entire decisions, in parallel.

Agent loops serialize control: a question, a model, an answer, and again. Kai is designed around a different primitive.

agent loop

question → model → answer
   ↓
question → model → answer
   ↓
question → model → answer
 
one call per decision, in series

decision programexample

faultbearing.91

severitycritical.86

readinessno-go.96

actioninspect-bearing.83

next_testspectrum.88

evidence_sufficienttrue.81

human_reviewyes.99

A Decision Program holds many typed variables, and Kai predicts their distributions together. One state. One shared encoding. Many decisions.

Refinement

Resolve certainty first.

Kai borrows from masked diffusion without pretending decisions are text tokens.

It starts from an incomplete decision state and predicts every unresolved variable in parallel. Confident nodes freeze. Only the uncertain or mutually dependent are refined.

Compute follows uncertainty, rather than revisiting everything.

refinementexample

start   [mask] [mask] [mask] … every node
 

100 decision nodes

pass 187 resolved pass 211 resolved constraint pass2 conflicts resolved

Dependencies

Dependencies are explicit.

Fully parallel does not mean independently wrong. A Decision Program is a graph.

Independent branches resolve together. Dependent branches resolve when the evidence they need is available.

The depth of inference follows the dependency graph, not the number of decisions. A ten-thousand-node program does not mean ten thousand serial model calls.

program graph

fault ─────┬──▶ action
           └──▶ readiness
 
severity ─────▶ human_review

Types

Decisions are typed.

Kai does not write a paragraph and leave software to work out what it meant.

choiceexample

Which subsystem is responsible?
 

bearing.91gearbox.05unknown.04

scoreexample

Operational risk
 

0.011.022.063.184.515.22

predicateexample

Is human approval required?
 

yes.97no.03

deferexample

When evidence does not support a safe decision,
not deciding is an answer.
 

decision_readyfalsereasoninsufficient evidencerequired_evidencevibration-spectrum

Scale

A million choices should still be a choice.

Flat classifiers degrade when every option has to fit inside one prompt. Kai separates discovery from judgment.

Candidate representations compile once with a Decision Program and are reused. That makes large tool catalogs, repositories, knowledge graphs and acquisition alternatives tractable, without showing every option to a frontier language model.

large choiceexample

1,000,000 candidates
   ↓ semantic retrieval
top 64
   ↓ kai reranking
top 8
   ↓ joint refinement
1

Specialists

One model should not pretend to be every model.

Kai is the decision layer. Specialists remain specialists.

A forecasting model forecasts. A vision model perceives. A solver satisfies hard constraints. A policy engine determines what is permitted. Kai decides across their evidence.

hanzo decision

hanzo decision
└─ kai          decision state
   ├─ zen       reasoning
   ├─ ts        forecasting
   ├─ graph     structure
   ├─ vision    perception
   ├─ audio     speech · signal
   └─ solver    math
 
policy          authority

Package

From perception to a Decision Package.

Consider a vehicle under evaluation.

vehicle-17 · evaluationexample

camera inspection
thermal video
radar
CAN telemetry
vibration
operator audio
maintenance history
SysML architecture
simulation results
requirements

decision programkaipolicy + solverhuman approvaldecision package

The result is not just an answer. It is a reproducible record of how the answer was reached.

Programs

Decision-as-a-program.

A Decision Program makes the structure explicit.

A program can combine deterministic calculations, Kai judgments, Zen reasoning, graph queries, simulations, forecasts, formal solvers and human approvals.

The model is part of the program. It is not the program.

power-system-selection.yamlexample

decision: power-system-selection
version: 7
objectives:
  - mission_range
  - reliability
  - maintainability
options:
  - architecture_a
  - architecture_b
  - architecture_c
constraints:
  max_mass: 480kg
  delivery: <= 18 months
assumptions:
  supplier_lead_time: 22 weeks
risks:
  - battery_availability
  - thermal_margin
approvals:
  - engineering
  - acquisition

Sensitivity

Ask what flips the decision.

A useful decision system tells you more than what it prefers. It tells you where that preference stops being true.

Kai and Hanzo Decision run sensitivity and counterfactual evaluation against the same versioned program. No new prompt. No hand-rewritten trade study.

power-system-selection · sensitivityexample

baselineArchitecture Abattery cost +11%A remains preferredbattery cost +17%A → Bmass limit −4%A remains preferredmass limit −8%A → C

Versions

Decisions that live longer than a meeting.

Evidence changes. Requirements change. Suppliers change. Models change. The Decision Program remains.

Hanzo records what evidence changed, which assumptions changed, which nodes recomputed, which boundary was crossed, why the result changed and who approved the new one.

A trade study becomes a decision system you can refresh.

power-system-selection · historyexample

version 14   Architecture A
   ↓ new reliability evidence
version 15   Architecture A
   ↓ mass requirement changed
version 16   Architecture B

Replay

Replay it.

Every Decision Package can preserve:

program version
input snapshots
evidence hashes
provenance
model revisions
Kai checkpoint
calibration revision
solver state
policy result
probability distributions
human overrides
approval state
execution trace

The question is not “What did the AI say?” It is “Can we reproduce the decision?”

Business

Autopilot for business.

Most of a business’s work is bounded decisions, followed by generation and execution.

Kai decides who, what, when and through which channel: enrichment, lead fit and intent, next action, offer, cadence, human handoff, and when to stop. It personalizes recommendations over the product graph.

Zen writes the message. Enso runs the workflow and routes each step to Kai, or to Zen when generation is the work or Kai defers. Policy decides whether contact is permitted: consent, opt-outs, caps, quiet hours and jurisdiction.

Configured natively in Hanzo Cloud, with signals from Hanzo Analytics, Insights and Commerce. Each program (marketing.*, sales.*, commerce.*) is measured by its own outcomes and tested in shadow before it acts.

Analytics Insights Commerce

sales.lead@4example

sales.lead@4 →

intent4fit5timingnownext_actiondemo

Agents

Built for agents.

The same primitive controls Hanzo’s own agent runtime.

Generative reasoning is reserved for the work that actually requires generation. Control becomes a typed system.

agent runtime · kai decides

which model?
which tool?
which file?
which skill?
how much context?
does this touch secrets?
should execution continue?
did the agent make progress?
does a human need to approve?

Physical systems

Built for physical systems.

Vehicles

Readiness, diagnostics, trade studies and maintenance.

Robotics

Perception-conditioned actions and safety escalation.

Manufacturing

Process state, anomalies, inspection and intervention.

Aerospace

Telemetry, digital engineering and configuration decisions.

Energy

Forecasting, equipment state and operational decisions.

Infrastructure

Sensors, simulations, risk and long-horizon planning.

One decision abstraction. Different evidence.

Deployment

Built to run where the evidence lives.

Decision evidence does not need to leave the infrastructure that owns it. Kai extends Hanzo’s operating plane with decision intelligence.

Hanzo Cloud
Your cloud
Your Kubernetes
Your GPU
On-prem
Disconnected environments

Results

Kai, against Laya and Jev.

The top mean accuracy across 11 application suites.

85.2%

Mean accuracy, 11 application suites

Jev 77.0% · Laya 70.2%

7 of 12

Suites won outright

DAIR Emotion, Banking77, Phishing, Toxicity, RAG relevance, Model routing, MASSIVE, 51 languages

7 · 1 · 4

Suites won, tied, behind

behind by points on AG News -0.7, Support triage -6.0, Jailbreak -2.0, Typed decisions -0.7

51 of 51

MASSIVE languages above Laya

macro accuracy 90.9% against 38.2%

Laya is the upstream kai-1 weights, served natively. Jev is typesafe/jev-1.13-20260917 through OpenRouter.

the frozen decision harness

	Accuracy			Calibration error (ECE)
Suite	Kai	Laya	Jev	Kai	Laya	Jev
AG News	94.3%	95.0%	86.0%	0.034	0.032	0.102
DAIR Emotion	93.5%	59.5%	60.3%	0.019	0.306	0.280
Banking77	90.8%	42.5%	83.5%	0.076	0.381	0.073
Support triage	44.3%	50.2%	36.5%	0.441	0.091	0.483
Email spam	99.8%	99.8%	97.8%	0.003	0.011	0.060
Phishing	99.0%	98.3%	90.0%	0.010	0.010	0.039
Jailbreak	92.0%	70.5%	94.0%	0.072	0.262	0.047
Toxicity	80.3%	53.0%	66.3%	0.191	0.297	0.177
RAG relevance	67.5%	62.5%	62.0%	0.029	0.117	0.279
Model routing	100.0%	63.9%	97.7%	0.000	0.089	0.014
Typed decisions	75.9%	76.6%	73.6%	0.172	0.213	0.047
MASSIVE, 51 languages	90.9%	38.2%	89.0%	0.060	0.418	0.064
Mean, 11 suites	85.2%	70.2%	77.0%	0.095	0.164	0.145

11 application suites and MASSIVE intent in 51 languages, the same questions to all three. Laya answers typed decisions on the checkpoint it ships for them.

Source: api.hanzo.ai/v1/research/runs?org=hanzo · harness · Kai a7@0834a74f · harness a1a4a73

Every language

massive intent, by language

	Accuracy
Language	Kai	Laya	Jev
Afrikaans	91.0%	34.0%	87.0%
Amharic	85.0%	15.0%	85.0%
Arabic	92.0%	46.0%	87.0%
Azerbaijani	94.0%	36.0%	88.0%
Bangla	92.0%	45.0%	86.0%
Welsh	86.0%	16.0%	76.0%
Danish	92.0%	46.0%	88.0%
German	91.0%	53.0%	93.9%
Greek	94.0%	44.0%	94.0%
English	94.0%	82.0%	92.0%
Spanish	92.0%	53.0%	93.0%
Persian	94.0%	51.0%	95.0%
Finnish	92.0%	30.0%	89.0%
French	91.0%	60.0%	90.0%
Hebrew	91.0%	37.0%	94.0%
Hindi	91.0%	46.0%	95.0%
Hungarian	90.0%	34.0%	92.0%
Armenian	89.0%	25.0%	84.0%
Indonesian	96.0%	36.0%	93.0%
Icelandic	85.0%	33.0%	84.0%
Italian	91.0%	47.0%	96.0%
Japanese	97.0%	64.0%	97.0%
Javanese	89.0%	16.0%	75.0%
Georgian	83.0%	15.0%	88.0%
Khmer	84.0%	20.0%	81.0%
Kannada	91.0%	30.0%	87.0%
Korean	91.0%	47.0%	94.0%
Latvian	89.0%	30.0%	85.0%
Malayalam	87.0%	28.0%	91.0%
Mongolian	88.0%	16.0%	84.0%
Malay	93.0%	27.0%	92.0%
Burmese	92.0%	16.0%	89.0%
Norwegian Bokmål	94.0%	50.0%	86.0%
Dutch	92.0%	40.0%	90.0%
Polish	94.0%	47.0%	92.0%
Portuguese	93.0%	44.0%	90.0%
Romanian	91.0%	37.0%	94.0%
Russian	97.0%	57.0%	94.0%
Slovenian	91.0%	33.0%	89.0%
Albanian	93.0%	29.0%	76.0%
Swedish	93.0%	48.0%	94.0%
Swahili	84.0%	13.0%	70.0%
Tamil	88.0%	31.0%	92.0%
Telugu	87.0%	22.0%	91.0%
Thai	93.0%	48.0%	94.0%
Filipino	91.0%	30.0%	85.0%
Turkish	90.0%	41.0%	92.0%
Urdu	90.0%	42.0%	90.0%
Vietnamese	91.0%	33.0%	89.0%
Chinese (China)	94.0%	65.0%	94.0%
Chinese (Taiwan)	95.0%	61.0%	93.0%

Twenty intents to choose from, a hundred utterances per language.

Source: api.hanzo.ai/v1/research/runs?org=hanzo · harness · Kai a7@0834a74f · harness a1a4a73

what a decision model can do

Measure	Kai	Laya	Jev
Choosing among many options
Accuracy, 4 options	–	77.8%	90.5%
Accuracy, 16 options	–	67.0%	85.3%
Accuracy, 77 options	–	42.5%	84.3%
Accuracy, 150 options	–	61.0%	94.0%
Accuracy, 1,000 options	–	0.0%	0.0%
Accuracy, 10,000 options	–	0.0%	0.0%
Accuracy, 100,000 options	–	0.0%	0.0%
Refused, 1,000 options	–	100.0%	100.0%
Refused, 10,000 options	–	100.0%	100.0%
Refused, 100,000 options	–	100.0%	100.0%
Five dependent questions on one state
Accuracy	–	33.6%	83.7%
All five right	–	0.5%	43.8%
Answers that break a rule	–	63.7%	24.8%
Accuracy, projected onto the rules	–	32.8%	87.9%
All five right, projected	–	4.0%	56.0%
Invariance
Answer changes when the options are reordered	–	25.5%	6.0%
Largest probability shift under reordering	–	0.996	0.470
A yes/no with its sides swapped: answer mirrors	–	20.5%	6.3%
A yes/no negated: answer mirrors	–	30.0%	88.3%
Sensor evidence, accuracy
Human activity, UCI HAR	–	20.3%	15.3%
Turbofan remaining life, C-MAPSS	–	17.0%	30.0%
Valve sound, MIMII	–	50.0%	84.0%
Conformal prediction sets at a 90% target
Coverage	–	90.1%	90.1%
Mean set size	–	6.801	1.304
Programs and deployment
Questions re-asked after one evidence change	–	100 of 100	100 of 100
Runs offline	–	yes	no
Self-hosted	–	yes	no
Evidence leaves the machine	–	no	yes
Pinned by	–	weights sha256	model name

The same cases to all three. Laya and Jev answer each question on its own; Kai refines dependent questions over three passes. Kai a7@0834a74f has no capability run yet, so no Kai figure shows here.

Source: api.hanzo.ai/v1/research/runs?org=hanzo · capability · Kai a7@0834a74f · harness a1a4a73

latency of one decision

Serving mode	Device	Dtype	p50 ms	p95 ms	p99 ms
Cold encode	ROCm	BF16	19.8	29.0	–
Incremental encode	–	–	–	–	–
Cached decision	–	–	–	–	–

Kai a7@0834a74f, one decision a call, timed on a production accelerator. A mode with no CUDA or ROCm measurement shows as absent; nothing here is projected.

Source: api.hanzo.ai/v1/research/runs?org=hanzo · latency · Kai a7@0834a74f · harness a1a4a73

Research

Research, not magic.

Kai is benchmarked against the systems it descends from and the ones it competes with, on the same questions, and every table above reproduces from the harness commit under it.

research tracks

01Extreme-cardinality decisionsFrom tens of candidates to hundreds of thousands.
02Multilingual decision intelligenceConsistent decisions across languages, rather than translation-first inference.
03Shared-state decodingEncode the evidence once and resolve many questions against it.
04Decision diffusionParallel structured inference with selective refinement.
05Multimodal evidenceSensors, video, audio, time series, graphs and engineering models.
06Calibration and abstentionKnowing when the evidence is insufficient.
07Decision graph consistencyMaking many locally plausible decisions globally coherent.

Read the specification, HIP-1332

Kai + Zen

They solve different problems.

generation · decision

zenkai

What should the implementation look like?Which implementation satisfies these constraints?

Explain the evidence.Is the evidence sufficient?

Generate a test plan.Which test should execute next?

Reason about the alternatives.Which decision state follows?

Generation and decision are complementary computational primitives. They should not be forced into the same model call.

The stack

Zen generates. Enso routes. Kai decides.

Policy governs. A model’s verdict can tighten a call and never loosen it.

API

Kai is exposed through Hanzo Decision.

One endpoint, POST /v1/decisions: a program, the evidence it reads, and typed answers back.

Kai returns a choice, not a completion. The example below is the shape of the call.

POST /v1/decisionsexample

POST /v1/decisions
{
  "model": "kai",
  "program": "vehicle.readiness@3",
  "evidence": [
    "sensor://vehicle-17/can",
    "video://inspection-382",
    "sysml://vehicle-17",
    "graph://maintenance-kg"
  ]
}
→
{
  "decision": "dec_01…",
  "answers": {
    "readiness": { "value": "no-go", "probability": 0.96 },
    "human_review": { "value": true, "probability": 0.99 }
  }
}

Open

Open research. Open runtime.

Kai builds on open research in typed decision models and evolves it into Hanzo’s native multimodal Decision runtime. Weights, benchmarks and runtime components are published as they reach reproducible release gates.

Run Kai through Hanzo Cloud. Run it on your own infrastructure. Study the benchmark. Fork the runtime.

Stop asking one model to do everything.

Reason with Zen. Perceive with specialist models. Forecast with temporal models. Verify with solvers. Govern with policy. Decide with Kai.

One decision plane over every kind of evidence.

Try Hanzo Explore Hanzo Decision