Try Hanzo
Hanzo · Enterprise

Kai (kai-1.8)

Decisions, not completions.

Shipping today

Model IDkai-1.8
Checkpointkai-1.8
Parameters~850M (multimodal decision architecture)
Context reach32,768 tokens (YaRN; trained: 1,024)
ModalityText and structured state
Output tokens0 (typed probability distribution)
EndpointPOST /v1/decisions
ArchitectureProprietary Decision Transformer + factorized cross-attention
Safety qualificationNot yet qualified: a sealed evaluation is being authored
Price…

In development

Kai-1.7 target
Parameters~850M (text, vision, audio, video)
Resident memoryUnder 2.0 GiB
Text latency<1 ms p50 (GPU), multi-pass adaptive refinement
Image latency360–400 ms (image + text state)
EncodersVision · Audio · Video
Reasoningreasoning: "auto" | "fast" | "deep"
TypeScript
// npm install hanzoai
import { AiApi, Configuration } from 'hanzoai'

const ai = new AiApi(new Configuration({ accessToken: process.env.HANZO_API_KEY }))

const { data: decision } = await ai.postDecisions({
  aiDecisionsRequest: {
    model: 'kai-1.8',
    state: { incident: 'checkout 5xx at 14%' },
    questions: {
      severity: {
        type: 'choice',
        instructions: 'How severe is this incident?',
        criteria: {
          low: 'cosmetic or internal; no customer is affected',
          high: 'some customers are blocked from finishing a task',
          critical: 'every customer is blocked and no revenue comes in',
        },
      },
    },
  },
})

const { choice, probabilities } = decision.answers.severity
console.log(choice, probabilities)
…. A native decision writes 0 output tokens; a decision Jev or a language model answers bills at that model’s rates, output included.

Limits

  • Reads text and structured state. Images, audio and video are the Kai-1.7 target.
  • Reaches 32,768 tokens by position; accuracy at that length is qualified for kai-1.4 and kai-1.5 only.
  • At most 100 questions a request. More questions, or a state past 32,768 tokens, is refused with 422.
  • The reasoning field is the Kai-1.7 target: the API refuses it today (400).

Use Kai (kai-1.8) via Hanzo AI

One API key. One call to POST /v1/decisions.

Build what’s next.