Kai-1.3: Production Decision Engine

⚠️ API-only · Weights not distributed
Kai-1.3 is Hanzo AI's production decision engine, hosted globally at POST /v1/decisions.
For local open-weight exploration, refer to hanzoai/kai-1-multilingual.

1. Model Overview

Kai-1.3 is a high-throughput, calibrated decision model designed to evaluate evidence and output typed decision states without generating natural language text. It replaces slow, non-deterministic autoregressive LLM classification with a specialized 382.3M parameter encoder architecture.

Primary Endpoint: https://api.hanzo.ai/v1/decisions
Identifier: kai or kai-1.3
Pricing: $0.021 / million input tokens (zero completion token surcharge)
Trained Context: 1,024 tokens (native encoder positions: 8,192; stretchable to 32,768 via YaRN 4x)
Serving Latency: ~12.4 ms p50 on standard CPU nodes

2. Architecture & Components

graph TD
    Context["Context & Evidence (Tokens)"] --> Encoder["Kai Encoder (22-Layer mmBERT-base)"]
    Encoder --> Latents["Latent Representations (768-dim)"]
    Options["Candidate Options (Invariant Batches)"] --> OptionEnc["Option Encoder (Cached)"]
    Latents --> Scorer["Dual-Block Cross-Attention Scorer"]
    OptionEnc --> Scorer
    Scorer --> Output["Calibrated Decision Probabilities & Typings"]

Encoder: 22-layer multilingual ModernBERT (mmBERT-base, hidden dim 768, 12 attention heads, unpadded rotary tables, 256,000 token vocabulary).
Dual-Block Cross-Attention Scorer: Evaluates options against state context with zero inter-option coupling (flip.order = 0.0000).
Decision Primitives:
- Bernoulli (boolean/noul): σ(s(true) − s(false))
- Choice: Softmax over top-K scored options (retrieval-narrowed up to 160 options).
- Score: Cumulative link / ranked probability scoring over discrete ordinal levels.

3. Verified Benchmark Results

All figures are reproducible via the official evaluation harness in github.com/hanzoai/decision/tree/main/bench:

Benchmark Suite	Metric	Kai-1.3	Jev 1.13	Baseline Laya	Standard LLM (GPT-4o)
Decision Index (Held-out)	Top-1 Accuracy	94.2%	89.1%	88.4%	91.2%
Decision Index	Brier Score (Calibration)	0.048	0.089	0.094	0.112
Decision Index	Expected Calibration Error	0.021	0.045	0.052	0.068
HelpSteer2	Preference Ranking Accuracy	88.6%	84.2%	83.1%	87.0%
Prompt Injection Suite	Attack Detection Recall	98.7%	91.5%	88.0%	94.2%
Order Invariance	Permutation Consistency	1.0000	0.9420	1.0000	0.8850
Inference Latency (p50)	CPU Wall Time	12.4 ms	480 ms	14.2 ms	850 ms

4. Intended Use & Workloads

Agent Routing & Tool Selection: Deciding which tool or API to call in autonomous workflows.
Guardrails & Policy Boundaries: Real-time content moderation, prompt injection detection, and deterministic policy enforcement.
Enterprise Triage: High-volume classification of customer inquiries, fraud indicators, and system anomalies.

5. Out-of-Scope Use

Freeform text generation, chat conversations, or essay writing (Kai outputs structured probabilities, not prose).
Autonomous life-critical decision systems without human verification.

6. Inference Code Snippets

cURL

curl -X POST https://api.hanzo.ai/v1/decisions \
  -H "Authorization: Bearer $HANZO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kai",
    "state": "User requested balance withdrawal of $15,000 from newly registered device.",
    "questions": [
      {
        "id": "risk_tier",
        "question": "What is the fraud risk tier?",
        "options": ["low", "medium", "high", "critical"]
      }
    ]
  }'

TypeScript SDK (`@hanzo/ai`)

import { HanzoAI } from '@hanzo/ai'

const hanzo = new HanzoAI({ apiKey: process.env.HANZO_API_KEY })

const decision = await hanzo.decisions.create({
  model: 'kai',
  state: 'Incoming webhook received with untrusted HMAC signature.',
  questions: [
    {
      id: 'webhook_action',
      question: 'Action required:',
      options: ['accept', 'quarantine', 'reject']
    }
  ]
})

console.log('Selection:', decision.answers.webhook_action.selection)
console.log('Confidence:', decision.answers.webhook_action.probabilities)

Python SDK (`hanzoai`)

from hanzoai import HanzoAI

client = HanzoAI()

res = client.decisions.create(
    model="kai",
    state="Pull request #124 modifies deployment config and opens port 22.",
    questions=[
        {
            "id": "requires_security_approval",
            "question": "Does this PR require security sign-off?",
            "options": ["yes", "no"]
        }
    ]
)
print(res.answers["requires_security_approval"].selection)

7. Cryptographic Verification & Hashes

Weights SHA-256: 0834a74f2d140642a453373da09e5a128e3d2c8d9e8bb1dcaf86d7210a4ecdfc
Representation Hash: 6dd51368d5b634cc6135da4984add5c15bf4b1e9977b4afbecff27ddac8d3777
Schema Hash: 9b83f624433c4cb5f01cae052440b15f5eb54392338d8ddbc1eefa0ae7cde5d7
Container Digest: v0.4.8@sha256:e495dd5f98a1a5b93c26d278e3aa122417b3cbc4c4ea0d2ef4902dffa545825a

API Documentation: docs.hanzo.ai/docs/decisions
Decision Recipes: docs.hanzo.ai/docs/guides/cookbook/recipes
Source Code: github.com/hanzoai/decision
Interactive Playground: huggingface.co/spaces/hanzoai/kai-playground

6.9 KiB Raw Permalink Blame History Unescape Escape