Inside Xiaomi Super XiaoAI: Fast/Slow Routing & Full-Duplex Dialogue for Automotive AI

Fast/Slow System: Intelligent Request Routing

A natural, fluid in-car conversation must be both fast and accurate. These goals often conflict: deeper reasoning improves accuracy but increases latency; skipping reasoning speeds responses but reduces correctness. Xiaomi's solution is a dual-system architecture inspired by human cognition.

Fast System handles high-frequency, simple commands — "turn on AC," "volume up," "navigate to office" — with low latency and low cost. Slow System handles complex, multi-step requests requiring deep reasoning: multi-waypoint navigation, conditional search, long-horizon task orchestration.

The router classifies each request by complexity and sends it to the optimal path. Example: "Navigate to office" (clear intent, single action) → Fast System, route planning starts immediately. "Find a campsite with Yangtze River view, free parking, and a mall within 2 km for dinner" (vague semantics, multiple constraints, tool chaining) → Slow System for semantic understanding, multi-condition filtering, and full itinerary planning.

Slow System Optimization: No-Think Mode + RAG-Planning

To make the Slow System both accurate and fast, Xiaomi upgraded the Xiaomi MiMo model in two stages:

Switch from think-mode to no-think-mode. The model stops "thinking aloud" and answers directly. This cut average model latency by ~35%, making XiaoAI feel "snappier." However, skipping reasoning hurt accuracy on complex tasks — the model would hallucinate or miss tool calls.

Introduce RAG-Planning. Before processing a user query, the system retrieves a few complete solution traces for similar problems from a knowledge base and feeds them as few-shot examples. The model then follows the demonstrated pattern instead of reasoning from scratch — an "open-book exam" approach. This restored think-mode accuracy while retaining no-think latency. End-to-end latency stays at no-think levels, 3–6 seconds faster than the previous think version, a perceptible difference in automotive scenarios.

Result: XiaoAI achieves no-think speed with think-mode accuracy.

Full-Duplex Dialogue: Listen-While-Speaking

Traditional voice assistants use half-duplex turn-taking: user finishes → assistant processes → assistant speaks → user speaks again. Super XiaoAI moves to full-duplex: the assistant listens continuously while speaking, enabling natural interruption and overlap.

A unified control module handles three decisions in real time: (1) Is the user addressing XiaoAI? (2) Has the user finished speaking? (3) Should the current playback be interrupted? This enables scenarios like: user says "Navigate to East Lake Luoyan Scenic Area," mid-sentence adds "Set AC to 24°C" — XiaoAI immediately executes the AC command, finishes the TTS, then resumes navigation planning.

Multi-Turn Context Compression (Three-Level Session Slimming)

Long trips generate dozens of dialogue turns; context tokens easily exceed 100k, degrading relevance and speed. Xiaomi designed a three-level compression pipeline:

Level 1 — Drop stale state. Per-second sensor snapshots (AC temperature, window position) are discarded by default.

Level 2 — Retain only current-task context. If navigating, prior music chat is pruned; overflow deletes oldest turns first, always preserving the most recent N turns.

Level 3 — Summarize ultra-long history. Beyond ~100 turns, history is condensed into a summary preserving long-term preferences.

This reduces token volume by 60–86% while maintaining or improving task completion and compliance — like a competent co-driver who remembers your destination but forgets irrelevant chatter.

Interruption Resumption Logic

Interruptions are classified into four types with distinct handling:

Broadcast interrupted (news, weather halfway): information already delivered → stop.

Task interrupted (navigation planning cut off): action incomplete → resume and finish.

False trigger (noise, cough): no real user intent → silently return to previous task.

Fast-path preemption (new quick command handled while background task runs): evaluate whether background task should continue.

Core criterion: does the interruption cause a missing action (resume) or only a missing information segment (stop)? The system also detects speech completeness using semantic coherence, prosody, and dialogue context, so it won't cut off a user who pauses mid-thought.

Deployment

Xiaomi SkyNomad (澎程) ships with the Expert Edition of Super XiaoAI. The same upgrades will roll out to other vehicle models and Xiaomi ecosystem devices.