查看購物車
  • English
  • Japanese
  • Korean
封面
市場調查報告書
商品編碼
2129077

汽車AI大規模模型技術調查報告，2026年

Automotive AI Large Model Technology Research Report, 2026
出版日期: | 出版商: ResearchInChina | 英文 390 Pages | 商品交期: 最快1-2個工作天內

價格
簡介目錄

汽車行業大規模人工智慧模型研究：競爭對手的關注點從「我們是否擁有更強大的模型？」轉變為「我們是否擁有更高的連結效率？」

ResearchInChina發布了《2026年汽車人工智慧大規模建模技術調查報告》。本報告概述了預計在2026年主導汽車產業的主流人工智慧大規模建模技術，深入分析了汽車人工智慧技術重點的轉變和發展趨勢。

2026年，汽車人工智慧大規模建模技術將主要應用於以下領域：

整體而言，與過去兩年大規模模式研發中以參數為主導的競爭相比，2026年的大規模模式技術發展更加重視實用性。在充分考慮技術創新的基礎上，技術發展更重視服務場景所需的技術路徑，並持續最佳化。同時，針對特定場景的技術，例如軌道規劃和基於提煉的部署，也開始呈現融合趨勢。儘管如此，在工程實施的細節方面，仍然湧現出獨特的解決方案和創新。簡而言之，2026年的大規模模型競爭從「誰擁有更強大的模型」轉向「誰擁有更高的鏈路級效率」。

此外，這些技術在應用場景中很少單獨實施。大多數情況下，它們會被組合起來建構解決方案。典型的組合包括多模態+知識圖譜、強化學習+擴散模型+特徵表示/建模最佳化、多智慧體系統+人工智慧編配以及訓練+配置。

以下將詳細說明一些選定的技術領域：

三種降低計算開銷的技術方法

儘管汽車晶片的運算能力不斷提升，但在駕駛座和自動駕駛領域並行執行複雜任務仍然會給運算能力帶來壓力。為了解決這個問題，在資源不變的情況下，可以透過三種技術手段來降低計算開銷：

模型端調整：參數縮放（例如，蒸餾、修剪、參數共用、低秩分解）、模型精度調整（例如，量化）以及模型架構的迭代改進（例如，注意力機制的變化、MoE）。

推理框架方面：預測解碼、KV快取最佳化、快取重複使用、提前終止等。

運算子和編譯器端特性：運算子融合、量化核心、顯式編配等。

調整模型參數和精確度是開銷最佳化中最常用的技術。與過去兩年相比，到2026年，OEM 廠商將能夠在模型架構和推理框架領域開發更豐富的技術，包括 MoE 和 KV 快取最佳化。

1.注意力機制的變體

2026年推出的注意力機制變體目的是應對汽車/邊緣設備運算能力受限的挑戰。它們解決了標準 softmax 注意力機制隨著序列長度增加而導致的計算和顯存開銷平方級成長的問題。此架構設計採用了「輕量級注意力機制變體 + 標準注意力機制」的組合。

在Xiaomi MiMo V2.5 系列的推理最佳化中，混合 SWA（滑動視窗注意力機制）將 KVCache 的儲存容量降低到全注意力機制的約 1/7。 MoE 利用稀疏活化來降低每個 token 的計算成本，同時維持模型容量。 KVCache 的減少不僅降低了計算開銷，還利用 SWA 最佳化釋放的顯存，將專家並行處理的大小減少到原來的一半，減少了機器間的通訊，將端到端預填充性能提升了約 40%，並提高了訓練期間的GPU 利用率。

2.教育部架構

MoE架構透過稀疏化控制推理階段的開銷，同時維持參數的完整性。2026年，OEM廠商的MoE技術將重點放在負載平衡、訓練和推理路由不匹配以及視覺標記冗餘。

即使採用相同的MoE 架構，路由器的處理方法也會因場景而異：

輸入導向活化（傳統模型驅動）：傳統模型驅動採用「輸入導向活化」，即活化哪個專家由當前輸入標記直接決定。這種方法在多任務和多模態輸入方面表現出色，但並不適用於複雜的多階段推理（例如圖形使用者介面代理）。

針對輸出的活化（CoME）：Xiaomi 的CoME 架構實現了「針對輸出的活化」，將專家活化與特定的推理階段（例如，螢幕摘要、子任務規劃、行動決策）聯繫起來。

場景特徵路由（EMoE）：BYD的自動駕駛解決方案採用了 EMoE。其路由器整合了一個獨立的場景路由器，可以直接接收場景特徵（例如，左轉、右轉、直行），並在所有層級上保持一致的路由選擇，加速模型收斂。

任務/模態感知路由：在Geely的解決方案中，MoE 專家調度單元透過門控網路分配權重，並根據當前任務類型（例如，記憶體儲存、記憶體搜尋）和模態特徵的權重，使用 Top 1 路由策略選擇主要專家。

3.修剪和量化

修剪和量化是部署模型的常用技術，尤其是在運算能力有限和延遲較高的設備上。

NAR的潛力正逐漸顯現。

目前，自回歸（AR）架構模型因其成熟的生態系統和明確的序列控制能力，已廣泛應用於語音互動、自動駕駛規劃、駕駛座代理等場景（例如 MoonCast、Transfuser 以及眾多基於 AR 的世界模型）。然而，由於 AR 模型以詞元/幀為單位進行序列生成，因此通常存在較高的推理延遲和穩定性風險（例如漏詞、難以理解的長片段、曝光偏差等）。

相較之下，非自回歸（NAR）模型近年來在多種場景下展現出許多優勢，包括更高的自適應穩定性、更高的效率和成本績效和駕駛規劃。Xiaomi ZipVoice-Dialog、OmniVoice、長安FlowR2A和Li Auto TrajHF等都是NAR模型的典型應用案例。儘管資料和部署範式等基礎支援在實用化應用場景有限，但NAR模型透過利用並行生成和流匹配/擴散範式，在多個基準測試中展現出超越AR基準模型的速度和品質，使其成為一種極具發展前景的技術方案。

NAR模型在生成速度、自適應後穩定性以及與計算資源的親和性表現出色，使其成為人工智慧驅動的生成和決策的新範式。儘管存在固有的自適應困難、資料支援挑戰以及工程路徑依賴性，但隨著技術的進步，NAR和AR模型有望長期共存。

多智慧體技術：整合調度核心+多樣化資訊共用機制

2026年至2027年間，模型供應商將繼續增強其底層模型的功能，而硬體供應商將以每年60%至70%的速度降低代幣成本。最大限度地利用大規模建模技術來應對當前挑戰是最重要的。汽車產業的一個關鍵例子是「始終線上智慧體」（Always on Agent），其主流應用屬於多智慧體技術範疇。

到2026年，多智慧體技術將實現技術路線的整合和應用場景的橫向擴展。它們採用主流的「中心+專家」模式，將應用範圍從孤立的駕駛座和自動駕駛場景擴展到涵蓋營運、維護、行銷和製造等跨系統、跨領域的協作場景。

每個原始設備製造商（OEM）都針對多智慧體系統實現了各自的智慧體協調機制。雖然集中式星型拓樸調度是應用最廣的，但在某些情況下，其他架構可能表現較佳。

2026年的多智慧體系統：

與其創建功能重疊的代理，不如為每個功能、角色和認知類型（規劃、探索、執行、防禦、分析、學習代理等）定義專用子代理，並讓它們執行各自的子任務。

該代理程式依靠各種協作機制（如標準化的通訊和狀態共用路徑，如 A2A 協定、全局狀態共用、資料互動介面、核心間指令幀、記憶體讀取/寫入操作和串流快取）來確保高效的資訊流和一致的狀態。

每個子代理完成任務後，中央模組會彙總並整合所有子任務的結果，輸出最優解並提供執行回饋。這實現了並行業務處理，提高了即時效能，並透過任務分區和冗餘備份增強了安全性和穩健性。

以一汽的基於代理的引擎ECU為例：

該解決方案在異構多核心 SoC 上部署了三個物理隔離的計算域，承載三種類型的代理：反應代理、規劃代理和學習代理。

該反應劑在符合 ASIL-D 標準的鎖定步進核心上運行，以亞 1 毫秒的周期收集感測器訊號，以毫秒級的延遲處理爆震和其他緊急運行條件，並在緊急情況下保持控制權。

規劃代理程式運作在符合 ASIL-B 標準的高效能核心上。它呼叫本地共用的數位孿生引擎（結合物理機制模型和神經網路）來模擬多個控制策略候選方案，透過多目標最佳化產生最佳控制參數，並透過結構化資料幀提供輸出，其中包括置信度、有效期和 CRC 校驗和。

學習代理程式運行在量子力學協處理器上。它在後台收集控制偏差資料，並且只對數位孿生模型執行增量更新。

這三個代理程式透過受 MPU 保護的共用記憶體分區和核心間中斷進行安全通訊，並輔以處理故障導致的效能下降和更新模型檢驗的機制。這種配置實現了完全封閉回路型、自最佳化的控制工作流程，其特點是「即時緊急應變- 針對未來的規劃 - 線上糾正」，同時符合 ISO 26262 的功能安全要求。

除了多智慧體技術之外，基於「Claw」的技術加速實現現役場景。

利安汽車的「StreamingClaw」就是一個已實現的代理範例：

這項技術利用了增量式流推理和主從代理機制。車輛不再將影片作為離線文件處理，而是像人類感知一樣，即時感知周圍環境並更新自身狀態。

「StreamingReasoning」負責即時感知和規劃，「StreamingMemory」透過分層記憶的演化累積多模態經驗，「StreamingProactivity」持續監控風險並觸發主動行動。

這使得車輛能夠實現一些功能，例如在駕駛員打哈欠或開車時使用行動電話時發出主動警告，在乘客上車時發出主動問候，以及即時識別乘客所持物品。

從車輛的角度來看，智慧體之間的競爭已經從「單一功能的競爭」演變為「涵蓋整個場景的生態系統體驗的競爭」。智慧體服務也從規範中列出的單一功能項轉變為與使用者積極互動，同時仍將駕駛主導委託給人類駕駛員的合作夥伴。

目錄

定義

第1章 大規模人工智慧模型應用技術的類型與趨勢

  • AI垃圾郵件模型的分類
  • 大規模人工智慧模型的定義和特徵
  • 基於架構的大規模人工智慧模型分類
  • 依任務類型和學習模式對大規模人工智慧模型進行分類；按監督模式對大規模人工智慧模型進行分類
  • 以模態對大規模人工智慧模型進行分類
  • 大規模人工智慧模型的應用場景
  • 基礎模型的分類
  • 汽車產業基礎模型的現狀
  • 此基礎模型在汽車領域的應用場景
  • 在駕駛座中使用大規模人工智慧模型的場景
  • 駕駛座代理使用場景分類（1）（2）
  • 大規模人工智慧模式在自動駕駛的應用場景
  • 大規模人工智慧模型在底盤領域的應用場景
  • 大規模人工智慧模式在動力傳動系統領域的應用場景
  • 大規模人工智慧模型在研究和設計領域的應用場景
  • 大規模人工智慧模型在製造業的應用場景
  • 在營運和行銷領域應用大規模人工智慧模型的場景
  • 人工智慧大規模模型技術的現狀
  • 感知技術：代表性分類
  • 語音/多模態對話技術：代表性分類
  • 感知技術（包括音訊和多模態）：技術路徑
  • 推理技術：代表性分類
  • 推理技巧：技術路徑
  • 學習與發展技巧
  • 學習與提煉技術：技術方法
  • 強化學習技術
  • 強化學習技術：技術路徑
  • 人工智慧大規模模式技術的發展趨勢
  • 人工智慧大規模模型技術的現狀
  • 感知技術的發展趨勢，包括音訊和多模態
  • 推理技術的發展趨勢（1）（2）
  • 學習與知識提煉科技的發展趨勢（1）（2）
  • 強化學習技術的發展趨勢：實現方案及相關技術

第2章 採用不同方法的大規模建模技術實例

  • 學習和最佳化技術
  • 利用大規模語言模型探索神經架構
  • 針對大規模模型的輕量級知識分發，將文字轉換為圖像
  • 輕量級視覺和語言模型（VLM）的知識蒸餾
  • 案例：訓練分類模型
  • 案例：學習軌道模型
  • 案例：強化學習的學習、決策與規劃模型
  • 案例：訓練人工智慧生成模型（從文字到圖像）
  • 案例：模型最佳化與測試
  • 案例：感知模型的迭代改進
  • 案例：最佳化融入情緒的問答產生模型的學習過程
  • 呼叫大規模建模工具
  • 推理技術
  • 推理範例：透過潛在空間推理提高Xiaomi 汽車的準確率
  • VLA模型的高效推理
  • 利用 MoE 架構加速大規模多模態視覺模型的推理
  • 大規模語言模型中的跨日曆時間推斷
  • 案例：加速模型推斷
  • 案例：邊緣雲協同推理
  • 案例：推理模型的最佳化
  • 邊緣部署技術
  • 案例：邊緣端模型推理的最佳化
  • 感知技術
  • 支援端對端自動駕駛的高保真虛擬仿真
  • 前瞻性預測世界建模
  • 案例：利用Xpeng實現預測推理
  • 加速用於自動駕駛的低步長自回歸視訊世界模型的推理
  • 案例：為Xpeng Auto的物理人工智慧系統「圖靈ViT」補充視覺功能
  • 4D3D場景佔用預測
  • 單目深度估計
  • 圖像和文字的精細跨模態對齊，以及圖像和文字搜尋
  • 多模態態對齊
  • 用於兩人之間自然對話的語音合成
  • 視訊時間定位多模態模型的訓練後最佳化
  • 最佳化多模態大規模模型視覺推理測試中的時間增強
  • 車內和車外的語音互動模式
  • 案例：利用多模態技術最佳化駕駛座功能操作
  • 案例：最佳化多模態互動
  • 案例：情緒辨識與舒適感的提供
  • 案例：情緒分析的多模態技術
  • 案例：結合口語對話和唇語
  • 案例：兒童監測中的多模態技術
  • 案例：下雨天照片拍攝
  • 案例：車聯網（IoV）的多模態感知
  • 代理互動技術
  • 案例：救援場景中的多智慧體協作
  • 案例：汽車製造商營運中的多智慧體協作
  • 案例：建立代理任務執行機制
  • 強化學習技術
  • 大規模混合專家（MoE）模式的強化學習微調
  • 大規模推理模型（LRM）搜尋能力下降
  • 案例：基於強化學習的學習系統（PPO）
  • 案例：使用強化學習產生資料
  • 知識圖譜技術
  • 知識圖譜範例：利用Xiaomi 汽車提高 RAG 準確率
  • 端對端音頻
  • 案例：利用Xiaomi 汽車最佳化語音模型
  • 案例：Xiaomi 汽車發表ASR/TTS整合車型
  • 案例：Xiaomi 汽車發布基於推理的外部語音場景語音模型
  • 案例：Xiaomi 汽車的語音模型提高了TTS的準確率
  • 智慧駕駛座中使用者語音指令的推理方法
  • 案例：基於強化學習的語音最佳化
  • 案例：提高語音辨識準確率
  • 案例：邊緣雲融合語音辨識

第3章 各種場景下的大規模建模技術實例

  • 世界模型
  • MiLA（多視角、高保真、長期影片生成）世界模型
  • 針對自動駕駛規劃模型的安全增強型學習最佳化
  • UniDrive WM：統一自動駕駛世界模型
  • 自動駕駛的基本世界渲染模型
  • 3D重建
  • 汽車資產3D模型的自動重建
  • 使用 3D 高斯散射進行重建
  • 車輛的高精度3D重建
  • 場景生成
  • 用於產生自動駕駛場景的擴散模型
  • 產生碰撞場景
  • 基於姿態的多角色動畫生成
  • 端對端自動駕駛
  • 僅基於視覺的端到端自動駕駛計劃
  • VLA端對端自動駕駛模型
  • VLA 時間因果推論與可解釋決策
  • 端對端自動駕駛
  • 僅使用視覺實現端對端自動駕駛
  • 營運風險預測
  • 端到端視覺和運動整合自動駕駛模型
  • 軌跡預測
  • 自動停車軌跡預測
  • TPK可靠軌跡預測模型
  • 自動駕駛軌跡預測
  • 周圍車輛的多模態未來軌跡預測
  • 軌跡規劃
  • 自動駕駛的多模態軌跡規劃
  • 自動駕駛的運動軌跡規劃
  • 在複雜的城市環境中規劃自動駕駛的軌跡
  • 車輛路線規劃
  • 自動駕駛車輛之間的自然語言通訊與協作
  • 自動駕駛中的賽道規劃
  • 案例：路徑規劃的多維評分策略
  • 案例：考慮時間窗的多智慧體路徑規劃
  • 案例：停車軌跡最佳化
  • 基準測試和評估
  • 評估多模態大規模自動駕駛模式在交通規則上的邏輯推理能力
  • 用於自動駕駛的多模態大規模交通模型
  • 推理可靠性評估
  • 自動駕駛中多條候選路線的評估
  • 可解釋人工智慧（xAI）的個人化適應
  • GUI代理
  • 全模態 GUI 代理的標準化評估
  • GUI視覺語言代理
  • 針對 GUI 自主代理的多階段混合推理
  • 積極互動
  • 代理驅動、主動式、服務導向的車輛控制
  • 為乘客提供智慧決策服務
  • 建立車輛人工智慧與使用者之間的雙向關係
  • 主動感知和互動，包括車內和車外
  • 基於類人記憶架構的人機介面（HMI）
  • 案例：駕駛座內的主動服務
  • 情感互動
  • 與機器人進行情感互動
  • 智慧情緒識別
  • 基於情感的智慧駕駛座控制方法
  • 用於駕駛員情緒辨識的自演化解決方案
  • 模組間控制
  • 駕駛座控制
  • 多端代理協同控制
  • 個人化車輛控制
  • 智慧駕駛座語音控制反應
  • 車輛控制
  • 智慧車輛使用提案生成方法
  • 全自動智慧後視鏡調整解決方案
  • 個人化視覺意圖識別解決方案
  • 案例：代理人對照明的調整
  • 案例：人工智慧驅動的座椅調節
  • 案例：人工智慧最佳化的自適應座椅調節
  • 案例：利用人工智慧調整座椅壓力分佈
  • 案例：零重力座椅的自適應調節
  • 案例：人工智慧在語音處理的應用
  • 案例：基於人工智慧的座艙罩控制
  • 行程服務
  • 駕駛座與駕駛員之間的整合多智慧體協調
  • 多智慧體合作
  • 加油推薦解決方案
  • 針對離線效能的智慧推薦方法
  • 案例：駕駛座內容推薦
  • 案例：地圖建構最佳化
  • 案例：害蟲防治
  • 內容創作
  • 語音驅動的影像生成
  • 自動產生圖畫書的方法
  • 案例：利用旅行人工智慧產生視訊BLOG
  • 個性化客製化
  • 自動產生汽車人機互動主題
  • 自適應汽車中央控制台介面產生解決方案
  • 為汽車產生虛擬化身
  • 案例：適應駕駛風格
  • 研究/設計
  • 車輛功能開發
  • 車輛模型部署框架
  • 案例：生產過程中的目視檢查
  • 行銷與營運
  • 用戶標籤生成
  • 車輛運算能力調度
  • 基於使用者資料的客製化車輛升級方法
  • 案例：雲端診斷
  • 動力傳動系統控制與最佳化
  • 案例：利用大規模模型的多參數融合技術進行電池熱失控預警
  • 案例：基於人工智慧的動態能源管理（1）（2）
  • 案例：基於代理技術的引擎控制
  • 案例：制定充電策略

第4章 人工智慧技術的未來應用

  • 案例：用於行銷評估評分的大規模建模技術
  • 案例：適應駕駛風格
  • 案例：複雜診斷工作條件下的代理技術
  • 案例：基於人工智慧的自適應多源功率調節技術
  • 案例：基於知識圖譜的意圖識別
  • 案例：駕駛座服務的多模型調度
  • 案例：應對幻覺
簡介目錄

Automotive AI Large Model Research: Competition Shifts from "Who Has the Stronger Model" to "Who Boasts Higher Link Efficiency"

ResearchInChina released the Automotive AI Large Model Technology Research Report, 2026. This report sorts out mainstream AI large model technologies in the automotive industry in 2026, and provides insights into shifts in technical focus of automotive AI and its trend evolution.

In 2026, automotive AI large model technologies are mainly deployed in the following fields:

Overall, compared with the race for parameter scale in large model development over the past two years, large model technologies in 2026 have grown more pragmatic. They focus vertically on the technical routes required by service scenarios for continuous optimization, while giving due consideration to innovation. Meanwhile, technologies tailored for specific scenarios, such as trajectory planning and distillation-based deployment, have begun to show signs of convergence. Nevertheless, distinctive solutions and innovations still emerge in the details of engineering implementation. In other words, the large model competition in 2026 is shifting from "Who has the stronger model" to "Who boasts higher link level efficiency".

Furthermore, these technologies are rarely deployed independently in application scenarios. They are frequently combined to build solutions. Typical combinations include multimodal + knowledge graph, reinforcement learning + diffusion model + feature representation & modeling optimization, multi agent system + AI orchestration, and training + deployment.

Several selected technical domains are elaborated below:

Three Technical Paths for Reducing Computational Overhead

Although in-vehicle chip computing power keeps growing, parallel execution of complex tasks in the cockpit and intelligent driving domains may still make computing power stretched. To address this problem, on the premise of unchanged resources, three technical paths exist for cutting computational overhead:

Model side: parameter scale adjustment (e.g., distillation, pruning, parameter sharing, low-rank decomposition), model precision adjustment (e.g., quantization), and model architecture iteration (e.g., attention mechanism variants, MoE).

Reasoning framework side: speculative decoding, KV cache optimization, cache reuse, early exiting, etc.

Operator and compiler side: operator fusion, quantized kernel, explicit orchestration, etc.

Adjusting model parameters and precision represents the most common overhead optimization route. Compared with prior two years, OEMs have richer technologies to develop in the model architecture and reasoning framework routes in 2026, including MoE and KV cache optimization.

1.Attention Mechanism Variants

In 2026 attention mechanism variants target sore points of constrained computing power on vehicle / edge devices. They tackle quadratic growth in computation and video memory overhead incurred by standard softmax attention as sequence length increases. Architecture design adopts a combination of "lightweight attention variants + standard attention".

In the case of reasoning optimization for Xiaomi's MiMo V2.5 series, Hybrid SWA (Sliding Window Attention) cuts KVCache storage to approximately 1/7 of that under Full Attention. MoE leverages sparse activation to hold model capacity while lowering per token computational cost. Reduced KV Cache not only cuts computation overhead but also makes use of video memory headroom from SWA optimization, trimming Expert Parallelism to 1/2 of its original scale, decreasing cross machine communication, lifting end to end Prefill performance by roughly 40%, and boosting training GPU utilization.

2.MoE Architecture

MoE architectures retain full scale parameters and control reasoning phase overhead via sparsification. In 2026, OEMs' MoE technologies focus on load balancing, training reasoning routing discrepancies, and visual token redundancy.

Even with the same MoE architecture, routers are handled differently for distinct scenarios:

Input-oriented activation (traditional MoE): Conventional MoE adopts "input-oriented activation", meaning activated experts are determined directly by current input tokens. This approach excels in multi task or multimodal input yet is suboptimal for complex multi-step reasoning (GUI Agents).

Output-oriented activation (CoME): Xiaomi's CoME architecture implements "output-oriented activation", aligning expert activation with specific reasoning phases (e.g., screen summarization, subtask planning, action decision making).

Scenario feature routing (EMoE): BYD's intelligent driving solution uses EMoE. Its router incorporates an independent scenario router that directly ingests scenario features (e.g., left turn, right turn, going straight), maintaining consistent routing selections across all layers to accelerate model convergence.

Task/modality aware routing: In Geely's solution, the MoE expert scheduling unit allocates weights via gating networks and selects leading experts with a Top 1 routing strategy, based on current task types (e.g., memory storage, memory retrieval) and modal feature weights.

3.Pruning & Quantization

Pruning and quantization are common techniques for model deployment under conditions of constrained computing power and increased latency, especially for on-device model deployment.

The potential of NAR begins to emerge.

At present, autoregressive (AR) architecture models, by virtue of mature ecosystems and explicit sequence supervision capabilities, are widely applied in voice dialogue, autonomous driving planning, cockpit agents and other scenarios (e.g., MoonCast, Transfuser and numerous AR based world models). Meanwhile, due to token by token/frame by frame sequential generation, AR models generally suffer from high reasoning latency and stability risks (e.g., word skipping, unintelligible long segments, and exposure bias).

In contrast, non-autoregressive (NAR) models have demonstrated a range of merits in recent years across scenarios including monolingual TTS, voice dialogue and driving planning: improved stability upon adaptation, higher efficiency and better cost performance. Typical examples include Xiaomi ZipVoice-Dialog, OmniVoice, Changan FlowR2A, and Li Auto TrajHF. Although their engineering deployment for complex tasks is still constrained by fundamental supports such as data and deployment paradigms, resulting in a limited scope of application scenarios, NAR models leverage parallel generation or flow-matching/diffusion paradigms to outperform AR baselines in both speed and quality on multiple benchmarks, making them one of the promising technical routes.

NAR models deliver advantages in generation speed, post adaptation stability and computing power friendliness, qualifying them as a new paradigm for AI generation and decision making. Restricted by native adaptation difficulty, data support and engineering path dependency, NAR and AR models are expected to co exist long term in technical evolution.

Multi-Agent Technology: Unified Scheduling Core + Diverse Information Sharing Mechanisms

During 2026 2027, model providers keep enhancing foundation model capabilities, while hardware vendors drag down token costs at an annual rate of 60-70%. How to fully leverage large model technologies to solve current challenges has become a top priority. On vehicles, one key embodiment is the "Always on Agent", whose mainstream construction modes falls under multi agent technology.

In 2026, multi agent technologies witness converging technical routes alongside horizontally expanded application scenarios. They adopt the mainstream "hub + expert" paradigm, extending from isolated cockpit and intelligent driving scenarios to cross system, cross domain collaborative scenarios including operation & maintenance, marketing and manufacturing.

Different OEMs implement distinct Agent collaboration mechanisms for multiagent systems. Centralized startopology scheduling sees the highest adoption frequency, yet alternative architectures yield superior outcomes for specific scenarios.

Multi-agent system in 2026:

Instead of creating agents with overlapping capabilities, dedicated sub-agents of function, role and cognition types (including planning, searching, execution, defense, analysis and learning agents) are defined to perform respective subtasks.

Agents rely on diverse collaboration mechanisms (standardized communication and state sharing paths such as A2A protocol, global state sharing, data interaction interfaces, inter-core instruction frames, Memory read-write operations, and streaming cache) to ensure efficient information flow and consistent states.

After each sub-agent completes its task, the central module aggregates all subtask outcomes, integrates and outputs the optimal solution, and feeds it back for execution. This enables parallel business processing to improve real-time performance, and enhances safety and robustness through task division and redundancy backup.

Take FAW's Agentbased engine ECU as an example:

This solution deploys three physicallyisolated computing domains on the heterogeneous multi-core SoC to host three types of Agents: Reaction Agent, Planning Agent and Learning Agent.

The Reaction Agent runs on ASIL-D lock-step cores, collecting sensor signals with <=1ms cycles, handling knock and other emergency operating conditions at millisecond-level latency and holding emergency takeover privileges.

The Planning Agent runs on ASIL-B high-performance cores. It invokes a local shared digital twin engine (combining physical mechanism models + neural networks) to simulate multiple candidate control strategies, generates optimal control parameters through multi-objective optimization, and delivers outputs via structured data frames carrying confidence levels, validity periods and CRC checksums.

The Learning Agent runs on QM coprocessors. It collects control deviation data in the background and only performs incremental updates for digital twin models.

These three agents communicate securely through MPU protected shared memory partitions and inter core interrupts, complemented by fault degradation and model verification update mechanisms. This setup delivers a full closed loop self-optimizing control workflow featuring "real time emergency response - forward looking planning - online correction" while complying with ISO 26262 functional safety requirements.

Beyond multi-agent technologies, active service scenarios see accelerated implementation with "Claw"-based technologies.

Li Auto's StreamingClaw serves as an example of an embodied agent:

It leverages incremental streaming reasoning plus master-slave proxies. Instead of processing video as offline files, vehicles perceive and update states in real time similar to human cognition.

StreamingReasoning handles real-time perception and planning; StreamingMemory stores multimodal experience through hierarchical memory evolution; StreamingProactivity continuously monitors risks and triggers proactive actions.

It enables functions such as proactive alerts when drivers yawn or operate mobile phones while driving, proactive greetings upon vehicle access, and real-time recognition of objects held by passengers.

From the vehicle perspective, agent competition evolves from "standalone feature battles" toward "fullscenario ecosystem experience battles". Agent services transform from discrete feature items on specification sheets into partners who proactively interact with users yet retain driving authority for human drivers.

Table of Contents

Definitions

1 Application Technology Types and Trends of AI Large Models

  • 1.1 Classification of AI Large Models
  • Definition and Characteristics of AI Large Models
  • Classification of AI Large Models by Architecture
  • Classification of AI Large Models by Task Type and Training Mode Classification of AI Large Models by Supervision Mode
  • Classification of AI Large Models by Modality
  • 1.2 Application Scenarios of AI Large Models
  • Classification of Foundation Models
  • Status Quo of Foundation Models in Automotive Industry
  • Application Scenarios of Foundation Models in Automobiles
  • Application Scenarios of AI Large Models in Cockpits
  • Classification of Cockpit Agent Application Scenarios (1)
  • Classification of Cockpit Agent Application Scenarios (2)
  • Application Scenarios of AI Large Models in Intelligent Driving
  • Application Scenarios of AI Large Models in Chassis Domain
  • Application Scenarios of AI Large Models in Powertrain Domain
  • Application Scenarios of AI Large Models in Research and Design
  • Application Scenarios of AI Large Models in Manufacturing
  • Application Scenarios of AI Large Models in Operation and Marketing
  • 1.3 Status Quo of AI Large Model Technologies
  • Perception Technology: Typical Classification
  • Voice/Multimodal Interaction Technology: Typical Classification
  • Perception Technology (Including Voice/Multimodal): Technical Path
  • Reasoning Technology: Typical Classification
  • Reasoning Technology: Technical Path
  • Training and Deployment Technology
  • Training/Distillation Technology: Technical Path
  • Reinforcement Learning Technology
  • Reinforcement Learning Technology: Technical Path
  • 1.4 Trends of AI Large Model Technologies
  • Current Layout of AI Large Model Technologies
  • Perception Technology Trends Including Voice/Multimodal (1): Vision Route
  • Perception Technology Trends Including Voice/Multimodal (2): Auditory Sense Route
  • Perception Technology Trends Including Voice/Multimodal (3): Implementation Solution and Supporting Technology
  • Reasoning Technology Trend (1): Intelligent Driving Scenarios
  • Reasoning Technology Trend (2): Cockpit Scenarios
  • Training and Distillation Technology Trend (1): Pre Training
  • Training and Distillation Technology Trend (2): Knowledge Distillation Evolves Toward Intra Policy Distillation
  • Reinforcement Learning Technology Trend: Implementation Solution and Grouped Technology

2 Large Model Technology Cases of Different Routes

  • 2.1 Training and Optimization Technology
  • Neural Architecture Search Driven by Large Language Model
  • Lightweight Knowledge Distillation for Text to Image Large Model
  • Knowledge Distillation for Lightweight Vision-Language Model (VLM)
  • Case: Classification Model Training
  • Case: Trajectory Model Training
  • Case: Training-Decision-Planning Model for Reinforcement Learning
  • Case: AI Generation Model Training (Text to Image)
  • Case: Model Optimization and Testing
  • Case: Perception Model Iteration
  • Case: Training Optimization for Emotional Q&A Generation Model
  • Large Model Tool Invocation
  • 2.2 Reasoning Technology
  • Reasoning Case: Xiaomi Auto's Latent Space Reasoning Improves Accuracy
  • Efficient Reasoning of VLA Model
  • Reasoning Acceleration for Multimodal Vision Large Model with MoE Architecture
  • Cross Calendar Time Reasoning of Large Language Model
  • Case: Model Reasoning Acceleration
  • Case: Edge-Cloud Collaborative Reasoning
  • Case: Reasoning Model Optimization
  • 2.3 Edge-side Deployment Technology
  • Case: Edge-side Model Reasoning Optimization
  • 2.4 Perception Technology
  • High Fidelity Virtual Simulation Supporting End-to-end Autonomous Driving
  • Forward-looking Predictive World Modeling
  • Case: Xpeng Realizes Forward-looking Reasoning
  • Reasoning Acceleration for Few-Step Autoregressive Video World Model for Autonomous Driving
  • Case: Visual Capability Supplement of Xpeng Auto's Physical AI System -TuringViT
  • 4D Three-Dimensional Scene Occupancy Prediction
  • Monocular Depth Estimation
  • Image-Text Fine-grained Cross-modal Alignment and Image-Text Retrieval
  • Multimodal Alignment
  • Voice Synthesis for Two-Person Natural Dialogue
  • Post-Training Optimization of Multimodal Model for Temporal Video Localization
  • Time Extension Optimization for Visual Reasoning Testing of Multimodal Large Model
  • In-vehicle and External Voice Interaction Modes
  • Case: Multimodal Technology Optimizes Cockpit Function Operation
  • Case: Multimodal Interaction Optimization
  • Case: Emotion Recognition and Soothing
  • Case: Multimodal Technology for Sentiment Analysis
  • Case: Lip Reading Combined with Voice Interaction
  • Case: Multimodal Technology for Child Monitoring
  • Case: Imaging in Rainy Conditions
  • Case: Multimodal Perception for Internet of Vehicles
  • 2.5 Agent Interaction Technology
  • Case: Multi-Agent Collaboration in Rescue Scenarios
  • Case: Multi-Agent Collaboration for Automaker Operation
  • Case: Establish Mechanism for Agent Task Execution
  • 2.6 Reinforcement Learning Technology
  • Reinforcement Learning Fine Tuning for Mixture of Experts (MoE) Large Model
  • Exploration Capability Collapse of Large Reasoning Model (LRM)
  • Case: Reinforcement Learning (PPO) Training System
  • Case: Data Generation via Reinforcement Learning
  • 2.7 Knowledge Graph Technology
  • Knowledge Graph Case: Xiaomi Auto Improves RAG Accuracy
  • 2.8 End-to-end Voice
  • Case: Voice Model Optimization by Xiaomi Auto
  • Case: Xiaomi Auto Launches ASR/TTS Single Model
  • Case: Xiaomi Auto Launches Reasoning Voice Model for External Voice Scenarios
  • Case: Xiaomi Auto's Voice Model Improves TTS Accuracy
  • Intelligent Cockpit User Voice Command Reasoning Method
  • Case: Voice Optimization via Reinforcement Learning
  • Case: Voice Accuracy Improvement
  • Case: Edge-Cloud Collaborative Voice Recognition

3 Large Model Technology Cases in Different Scenarios

  • 3.1 World Model
  • MiLA Multi-view Intensive-fidelity Long-term Video Generation World Model
  • Safety Reinforcement Learning Optimization for Autonomous Driving Planning Model
  • UniDrive WM Unified Autonomous Driving World Model
  • Basic World Rendering Model for Autonomous Driving
  • 3.2 3D Reconstruction
  • Automatic Vehicle 3D Asset Reconstruction
  • 3D Gaussian Splatting Reconstruction
  • Vehicle High Precision 3D Reconstruction
  • 3.3 Scene Generation
  • Diffusion Model for Autonomous Driving Scene Generation
  • Collision Scene Generation
  • Pose-Driven Multi-Character Animation Generation
  • 3.4 End-to-end Autonomous Driving
  • Vision-only End-to-end Autonomous Driving Planning
  • VLA End-to-end Autonomous Driving Model
  • VLA Temporal Causal Reasoning and Interpretable Decision
  • End-to-end Autonomous Driving
  • Vision-only End-to-end Autonomous Driving
  • Motion Risk Prediction
  • End-to-end Vision-Action Autonomous Driving Model
  • 3.5 Trajectory Prediction
  • Automatic Parking Trajectory Prediction
  • TPK Trusted Trajectory Prediction Model
  • Autonomous Driving Trajectory Prediction
  • Multimodal Future Trajectory Prediction of Surrounding Vehicles
  • 3.6 Trajectory Planning
  • Multimodal Trajectory Planning for Autonomous Driving
  • Motion Trajectory Planning for Autonomous Driving
  • Autonomous Driving Trajectory Planning in Complex Urban Environment
  • Vehicle Path Planning
  • Natural Language Communication and Collaboration between Autonomous Vehicles
  • Autonomous Driving Trajectory Planning
  • Case: Multi-Dimension Scoring Strategy for Path Planning
  • Case: Multi-Agent Path Planning with Time Window
  • Case: Parking Trajectory Optimization
  • 3.7 Benchmark Test and Evaluation
  • Evaluation on Traffic Rule Logical Reasoning Capability of Autonomous Driving Multimodal Large Model
  • Autonomous Driving Multimodal Large Model Traffic
  • Chain of Thought Credibility Evaluation
  • Multi-Candidate Trajectory Evaluation for Autonomous Driving
  • Personalized Adaptation of Explainable AI (xAI)
  • 3.8 GUI Agent
  • Standardized Evaluation for Omnimodal GUI Agent
  • GUI Vision-Language Agent
  • Multi-Stage Hybrid Reasoning for GUI Autonomous Agent
  • 3.9 Proactive Interaction
  • Agent Proactive Service-Oriented Vehicle Control
  • Passenger Intelligent Decision Service
  • Two-Way Cultivation between Vehicle AI and Users
  • In/Out-cabin Proactive Perception Interaction
  • HMI Based on Human-like Memory Architecture
  • Case: Cockpit Proactive Service
  • 3.10 Emotional Interaction
  • Robot Emotional Interaction
  • Intelligent Emotion Recognition
  • Intelligent Cockpit Control Method Based on Emotion Recognition
  • Self-Evolution Solution for Driver Emotion Recognition
  • 3.11 Cross-Module Control
  • Cockpit Control
  • Multi-End Agent Collaborative Control
  • Personalized Vehicle Control
  • Voice Control Response of Intelligent Cockpit
  • Vehicle Control
  • Intelligent Vehicle Usage Suggestion Generation Method
  • Fully Automatic Intelligent Adjustment Solution for Rearview Mirror
  • Personalized Visual Intention Recognition Solution
  • Case: Light Adjustment via Agent
  • Case: Seat Adjustment by AI
  • Case: Adaptive Seat Adjustment Optimized by AI
  • Case: Seat Pressure Distribution Adjustment by AI
  • Case: Adaptive Adjustment for Zero Gravity Seat
  • Case: AI Applied to Audio Processing
  • Case: Canopy Control by AI
  • 3.12 Mobility Service
  • Cockpit-Driving Integrated Multi-Agent Collaboration
  • Multi-Agent Cooperation
  • Refueling Recommendation Solution
  • Intelligent Recommendation Method for Offline Performance
  • Case: Cockpit Content Recommendation
  • Case: Map Construction Optimization
  • Case: Living Creature Repellent
  • 3.13 Content Creation
  • Voice-Driven Image Generation
  • Automatic Picture Book Generation Method
  • Case: Travel AI Vlog Generation
  • 3.14 Personalized Customization
  • Automatic Generation of Automotive HMI Theme
  • Adaptive Automotive Center Console Interface Generation Solution
  • Automotive Virtual Avatar Generation
  • Case: Driving Style Adaptation
  • 3.15 Research and Design
  • Vehicle Function Development
  • Vehicle Model Deployment Framework
  • Case: Visual Inspection in Production
  • 3.16 Marketing and Operation
  • User Tag Generation
  • Vehicle Computing Power Scheduling
  • Customized Vehicle Upgrade Method Driven by User Portrait
  • Case: Cloud Diagnosis
  • 3.17 Powertrain Control and Optimization
  • Case: Large Model Multi-Parameter Fusion Technology for Battery Thermal Runaway Warning
  • Case: AI-enabled Dynamic Energy Management (1)
  • Case: AI-enabled Dynamic Energy Management (2)
  • Case: Engine Control via Agent Technology
  • Case: Charging Strategy Decision

4 Application of Forward-looking AI Technologies

  • Case: Large Model Technology for Marketing Evaluation Scoring
  • Case: Driving Style Adaptation
  • Case: Agent Technology for Complex Diagnostic Working Condition
  • Case: AI Adaptive Multi-Source Power Coordination Technology
  • Case: Intention Recognition Based on Knowledge Graph
  • Case: Multi-Model Scheduling of Cockpit Services
  • Case: Hallucination Handling