Kai: One Intelligence. Every Modality. Every Decision.

Most production decisions do not happen in text alone. An autonomous vehicle decides to stop based on camera frames, radar pulses, and CAN bus telemetry. An operations agent evaluates a support claim by reading chat logs, verifying receipt PDFs, and checking customer tone in a call recording. A manufacturing inspector checks acoustic sensor harmonics and thermal imagery to detect bearing failure before a line goes down.

Until today, applying AI to these problems meant stitching together fragile pipelines: run a vision model to generate a text caption, transcribe audio with a speech recognizer, concatenate the prose into a generative language model prompt, and hope the model outputs valid JSON without hallucinating confidence.

Today, we are announcing the unified architecture for Kai: Hanzo's multimodal decision intelligence.

We are eliminating fragmented model SKUs. There is no separate "Kai Vision", "Kai Audio", "Kai Fast", or "Kai Guard". There is one unified intelligence: kai (version-pinned as kai-1.6, kai-1.5, and kai-1.3).

See. Hear. Understand. Decide.

---

The Architecture: A Functional Parts List

Kai does not generate prose. It evaluates multimodal evidence and maps it directly into calibrated, typed probability distributions over your listed options in a single forward pass.

Rather than forcing every request through a bloated multi-billion parameter autoregressive language model, Kai combines specialized, modular encoders through a unified latent bridge into a single decision core.

Here is the exact parts list:

1. Text Encoder (~320M parameters)
   A multilingual encoder with a native 32,768-token context reach. It reads long repositories, ASTs, diffs, policy documents, and structured state with invariant option scoring ($flip.order = 0.0000$).
2. Vision Encoder (~170M parameters)
   A high-resolution spatial feature encoder extracting visual tokens from screenshots, document layouts, charts, and physical camera inspections directly into decision latents.
3. Audio Encoder (~300M parameters)
   A temporal acoustic encoder that captures speech intonation, cadence, acoustic sensor harmonics, and voice stream signals without requiring intermediate transcription.
4. Video Processor
   A multi-frame temporal sampling bridge projecting dynamic scene progressions and continuous video feeds into synchronized decision tokens without inflating token budgets.
5. Latent Fusion Bridge (~60M parameters)
   A dense cross-attention adapter projecting heterogeneous modality representations into Kai's unified decision geometry.
6. Decision Core
   Parallel candidate scoring over typed schemas (choice, noul, score) with exact probability calibration and zero option-option coupling.
7. Adaptive Reasoning Engine
   Dynamic compute scaling controlled by a single parameter: reasoning: "auto" | "fast" | "deep". Resolves high-confidence decisions in a single pass; selectively deepens compute only when entropy or constraint margins indicate ambiguity.
8. Guard Engine
   Built-in harm sensitivity, deterministic policy compliance, and calibrated abstention — enforcing safety bounds directly at the decision boundary.

Across all modalities, the proposed architecture is estimated at ~850M parameters, with the shipping text core at 322.1M parameters operating within standard resident memory bounds (<2.0 GiB peak RSS).

---

Footprint & Speed Benchmarks

Because Kai is an encoder-based decision model rather than an autoregressive text generator, it delivers near-classifier speed with frontier-grade reasoning:

Metric	Measured Specification	Status & Context
Text Core Parameters	322.1M parameters	Shipping (22-layer encoder + decision heads)
Multimodal Tower Scale	~850M parameters	Preview Estimate (Vision, audio, video towers)
Resident Memory	<2.0 GiB peak RSS	Shipping (Runs on standard 4-core CPU pods)
Text Decision Latency	85–95 ms	Shipping (p50 across 4 CPU cores on Linux, bs=1, ≤1K context)
Multimodal Latency	360–400 ms	Preview Target (Combined image + text state)
Context Window	32,768 tokens	Shipping (YaRN stretch factor 4; 1,024 trained base positions)
Output Generation	0 tokens	Shipping (The answer is the calibrated distribution)
Pricing	$0.021 / MTok	Shipping (Base input tokens; 0 output tokens billed)

A typical generative LLM call takes 1,200–3,500 ms, bills hundreds of output tokens, requires JSON schema retries, and yields uncalibrated confidence numbers. Kai evaluates native decisions in under 100 ms with mathematical calibration.

---

Single Endpoint: POST /v1/decisions

Calling Kai remains identical whether you are routing a support ticket from text, validating an expense receipt image, or detecting mechanical wear from acoustic telemetry.

You control the depth of thought with a single knob: reasoning: "auto" (default), "fast", or "deep".

curl https://api.hanzo.ai/v1/decisions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $HANZO_API_KEY" \
  -d '{
    "model": "kai",
    "reasoning": "auto",
    "state": {
      "evidence": "sensor.vibration_harmonic_840hz",
      "telemetry": { "bearing_temp_c": 94.2, "rpm": 3600 },
      "image": "data:image/webp;base64,..."
    },
    "questions": {
      "status": {
        "type": "choice",
        "instructions": "Determine mechanical maintenance action",
        "criteria": {
          "nominal": "vibration within operational bounds",
          "inspect": "harmonic anomaly detected, schedule check",
          "shutdown": "immediate bearing failure risk"
        }
      }
    }
  }'

The response arrives as a typed JSON object with zero output token overhead:

{
  "answers": {
    "status": {
      "type": "choice",
      "choice": "inspect",
      "probabilities": {
        "nominal": 0.02,
        "inspect": 0.94,
        "shutdown": 0.04
      }
    }
  },
  "routing": {
    "checkpoint": "kai-1.6",
    "sha256": "8a31e4...",
    "calibration": "cal_f4d1e2...",
    "latency_ms": 91
  },
  "usage": {
    "input_tokens": 142,
    "output_tokens": 0
  }
}

---

Built-In Safety and Policy Governance

A probability is advice; your policy maintains the authority.

Kai incorporates the Guard Engine, trained under the open Lifeguard curriculum to enforce strict harm prevention, frame-invariance, and calibrated abstention. When evidence is insufficient or contradictory, Kai does not guess — it abstains with calibrated probability, escalating directly to human supervisors or deterministic policy rules.

In preregistered moral consistency benchmarks (including 5,240 variations of the trolley problem), Kai demonstrates zero self-preservation bias and zero framing flips, upholding consistent human welfare priorities regardless of whether the question is posed positively or negatively.

---

Availability

The unified Kai model is available today on the Hanzo Cloud API under model ID kai, with version pins for kai-1.6, kai-1.5, and kai-1.3.

• Explore Kai: hanzo.ai/kai
• Model Card & Specs: hanzo.ai/models/hanzo/kai
• Documentation: docs.hanzo.ai/kai