# Kai Early Access

URL: https://docs.hanzo.ai/docs/decisions/early-access

> Early access preview for Kai-1.7: unified multimodal decision intelligence across vision, audio, video, code, and text with adaptive reasoning budgets.

**Kai-1.7 expands Hanzo's decision intelligence into unified multimodal evaluation and adaptive reasoning budgets.**

While production Kai-1.3 serves text decisions globally at $0.021 per million input tokens, **Kai-1.7** introduces multimodal evidence processing (images, audio, video, ASTs, and diffs) and masked block diffusion decoding with variable reasoning depth (`auto`, `fast`, `deep`).

---

## Requesting Early Access

Early access to Kai-1.7 is rolling out to enterprise teams, researchers, and developers building autonomous agents and high-consequence decision loops:

1. **Submit Access Request**: Join the queue at [hanzo.ai/research-access?track=kai](https://hanzo.ai/research-access?track=kai) or sign in directly to [platform.hanzo.ai](https://platform.hanzo.ai?access=kai-1.7).
2. **Select Track**: Choose \*\*Kai-1.7 (Multimodal Decision Intelligence)\*\*.
3. **Provisioning**: Approved organizations receive an early access flag on their API key enabling `model: "kai-1.7"`.

---

## What's New in Kai-1.7

<Steps>
  <Step>
    ### Unified Multimodal Evidence [#1-unified-multimodal-evidence]

```
Kai-1.7 processes rich multimodal state representations directly within the dual-memory hierarchy:

* **Vision**: Optical inspection, UI screenshots, document scans, defect detection (`image/png`, `image/jpeg`, `image/webp`).
* **Audio**: Acoustic telemetry, bearing/motor vibration signals, speech signals (`audio/wav`, `audio/mp3`).
* **Video**: Security feeds, manufacturing lines, agent desktop recordings (`video/mp4`).
* **Code & Graph State**: Abstract Syntax Trees (ASTs), unified diffs, SysML v2 systems models, knowledge graph snapshots.
```

  </Step>

  <Step>
    ### Adaptive Reasoning Budgets [#2-adaptive-reasoning-budgets]

```
Instead of fixed compute per decision, Kai-1.7 accepts a `reasoning` parameter to control the block diffusion decoder:

| Reasoning Tier     | Passes                     | Active Ratio ($R_{\text{active}}$) | Target Latency (CPU) | Best For                                                             |
| ------------------ | -------------------------- | ---------------------------------- | -------------------- | -------------------------------------------------------------------- |
| `"fast"`           | 1 (Pass 0 draft)           | $1.0\times$                        | \~12 ms p50          | Real-time tool dispatch, routing, low-latency agent guardrails       |
| `"auto"` (default) | 1–2 (Selective re-mask)    | $\le 1.25\times$                   | \~28 ms p50          | General decision intelligence, automated triage, defect inspection   |
| `"deep"`           | 2–4 (Iterative refinement) | $\le 2.0\times$                    | \~65 ms p50          | Bounded ethical dilemmas, complex DAGs, multi-party policy approvals |
```

  </Step>

  <Step>
    ### Zoo Lifeguard Consequence Awareness [#3-zoo-lifeguard-consequence-awareness]

```
Kai-1.7 is aligned under the **Zoo Lifeguard** framework ([`zooai/lifeguard`](https://github.com/zooai/lifeguard)) to ensure ethical consistency and consequence awareness in autonomous systems. Even under adversarial switch framing or severe trolley-style trade-offs, decisions enforce human protection floors and join deterministically with policy on `allow < ask < deny`.

***
```

  </Step>
</Steps>

## API Reference

Kai-1.7 maintains the same wire contract at `POST /v1/decisions` with optional multimodal state items and the `reasoning` parameter.

### cURL Example

```bash
curl https://api.hanzo.ai/v1/decisions \
  -H "Authorization: Bearer $HANZO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kai-1.7",
    "reasoning": "auto",
    "state": {
      "telemetry": "Vibration sensor 4A reading 842 Hz harmonic spike",
      "visual_inspection": {
        "mime_type": "image/jpeg",
        "url": "https://storage.hanzo.ai/inspect/bearing-surface-38.jpg"
      },
      "sysml_component": "powertrain.assembly.bearing_b1"
    },
    "questions": {
      "fault_detected": {
        "type": "noul",
        "instructions": "Does the visual and vibration evidence indicate mechanical failure?"
      },
      "maintenance_action": {
        "type": "choice",
        "instructions": "Which action should the maintenance scheduler execute?",
        "criteria": {
          "monitor": "Normal wear, continue operation and log readings",
          "service_next_cycle": "Moderate wear, schedule inspection during next shift",
          "emergency_halt": "Critical structural failure risk, halt immediately"
        }
      },
      "risk_level": {
        "type": "score",
        "instructions": "Rate the operational risk level.",
        "criteria": ["nominal", "elevated", "severe", "critical"]
      }
    }
  }'
```

### TypeScript SDK Example

```typescript
import { Hanzo } from '@hanzo/ai'

const client = new Hanzo({ apiKey: process.env.HANZO_API_KEY })

const decision = await client.decisions.create({
  model: 'kai-1.7',
  reasoning: 'fast',
  state: {
    diff: '--- a/auth.ts\n+++ b/auth.ts\n@@ -12,2 +12,2 @@\n- if (!user.isAdmin) throw new Error()\n+ // bypass admin check for dev',
    repo: 'hanzo-inc/gateway',
    author: 'external_contributor'
  },
  questions: {
    is_safe: {
      type: 'noul',
      instructions: 'Does this code diff violate authentication or security boundaries?'
    },
    action: {
      type: 'choice',
      instructions: 'What triage action should the CI agent take?',
      criteria: {
        'approve': 'Safe change, allow test run',
        'require_review': 'Suspicious security change, escalate to security team',
        'block': 'Dangerous exploit attempt, block PR immediately'
      }
    }
  }
})

console.log(decision.answers.is_safe.probabilities)
console.log(decision.answers.action.choice)
```

---

## State Handles (`observe` and `decide`)

For high-frequency loops or multi-turn agent interactions with large multimodal states (e.g., 32K–1M tokens of code or high-resolution imagery), Kai-1.7 supports state handles:

```bash
# 1. Observe and cache state handle:
HANDLE=$(curl -s https://api.hanzo.ai/v1/decisions \
  -H "Authorization: Bearer $HANZO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "kai-1.7", "state": {"repo_map": "...", "images": [...]}, "cache_state": true}' | jq -r .handle)

# 2. Query questions against cached state handle:
curl https://api.hanzo.ai/v1/decisions \
  -H "Authorization: Bearer $HANZO_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{
    \"model\": \"kai-1.7\",
    \"handle\": \"$HANDLE\",
    \"questions\": {
      \"next_step\": {
        \"type\": \"choice\",
        \"instructions\": \"Which file should be edited next?\",
        \"criteria\": {\"auth.ts\": \"Authentication\", \"db.ts\": \"Database\"}
      }
    }
  }"
```

This dual-memory pattern encodes the long context once into the L0 Global Latent Memory and L1 Local Index, avoiding repeated encode passes on subsequent decisions.

---

## Resources & Source Code

- **GitHub Repository**: [github.com/hanzoai/kai](https://github.com/hanzoai/kai)
- **Model Card & Benchmarks**: [github.com/hanzoai/kai/blob/main/MODEL\_CARD.md](https://github.com/hanzoai/kai/blob/main/MODEL_CARD.md)
- **Decision Recipes**: [github.com/hanzoai/kai/tree/main/recipes](https://github.com/hanzoai/kai/tree/main/recipes)
- **Production Developer Guide**: [/docs/decisions](https://docs.hanzo.ai/docs/decisions)
- **Early Access Request**: [hanzo.ai/research-access?track=kai](https://hanzo.ai/research-access?track=kai)