ZoZo Auto Research has released the "2026 Automotive AI Large Model Technology Research Report", providing insights into the shifting focus and evolving trends of automotive AI technologies by reviewing the mainstream AI large model technologies in the automotive industry in 2026.
In 2026, automotive AI large model technologies are primarily deployed in the following areas:
Figure: Mainstream Automotive AI Large Model Technology Types (Partial)
Compiled by: ZoZo Auto Research
Table: Mainstream Automotive AI Large Model Technology Cases (Partial)
Compiled by: ZoZo Auto Research
Overall, compared to the parameter scale competition of the past two years, large model technologies in 2026 are more pragmatic. They focus vertically on the technical routes required by service scenarios, continuously optimizing while balancing innovation. Technologies serving specific scenarios (such as trajectory planning and distillation deployment) have begun to show signs of convergence, yet they still demonstrate unique solutions and innovative points in the details of engineering implementation. In other words, the competition in large model technologies in 2026 is shifting from "who has the stronger model" to "who has the higher pipeline efficiency".
Meanwhile, the above technologies are not used independently in application scenarios; they are often mixed and matched with other technologies to build solutions. Typical examples include multimodal + knowledge graph, reinforcement learning + diffusion model + feature representation and modeling optimization, multi-agent system + AI orchestration, and training + deployment.
Mix and Match of Different Technical Routes in Specific Scenarios
Compiled by: ZoZo Auto Research
The following elaborates on several specific technical fields:
Three Technical Paths to Reduce Computational Overhead
The computing power of in-vehicle chips continues to increase, but the parallel execution of complex tasks in the cockpit and intelligent driving domains may still strain the computing resources. To address this issue, under the premise of unchanged resources, there are three technical paths from the perspective of reducing computational overhead:
Model side: Parameter scale adjustment (e.g., distillation, pruning, parameter sharing, low-rank factorization, etc.), model precision adjustment (e.g., quantization, etc.), and model architecture iteration (e.g., attention mechanism variants, MoE, etc.).
Inference framework side: Speculative decoding, KV cache optimization, cache reuse, early exit, etc.
Operator and compiler side: Operator fusion, quantized kernels, explicit orchestration, etc.
Among them, the most commonly used overhead optimization route is adjusting model parameters and precision. Compared to the past two years, in 2026, automakers' technical development in routes such as model architecture and inference frameworks has also become richer, including technical means such as MoE and KV cache optimization.
01. Attention Mechanism Variants
In 2026, attention mechanism variants primarily address the pain points of "in-vehicle/edge-side constrained computing power" scenarios, focusing on solving the computational and memory overhead of standard softmax attention that grows quadratically with sequence length. In terms of architecture design, a combination of "lightweight attention variant + standard attention" is adopted:
Usage of Attention Mechanism Variants by Some Automakers
Compiled by: ZoZo Auto Research
Taking Xiaomi's MiMo-V2.5 series inference optimization process as an example, adopting Hybrid SWA (Sliding Window Attention) reduces KVCache storage to about 1/7 of Full Attention; MoE reduces the computational cost per token while maintaining model capacity through sparse activation. The reduced KVCache not only lowers computational overhead but also utilizes the VRAM margin optimized by SWA to reduce Expert Parallelism to 1/2 of the original, decreasing cross-machine communication and improving end-to-end Prefill performance by about 40%, thereby enhancing the utilization rate of training cards.
Xiaomi Applies the SWA Algorithm in the Overhead Optimization Process
Image source: Xiaomi MiMo
Key Points of Inference Optimization for the MiMo-V2.5 Series
Data source: Xiaomi MiMo; Compiled by: ZoZo Auto Research
02. MoE Architecture
The MoE architecture does not cut parameters but controls the overhead of the inference process through sparsification. In 2026, automakers' MoE technologies focus on issues such as load balancing, training-inference routing discrepancies, and visual token redundancy.
Usage of MoE Technology
Data source: Public information; Compiled by: ZoZo Auto Research
Even when adopting the same MoE architecture, the router processing methods vary across different scenarios:
Input-oriented activation (Traditional MoE): Traditional MoE relies on "input-oriented activation," which directly determines which expert to activate based on the current input token. This approach has advantages in handling multi-task or multimodal inputs but is not optimal in scenarios requiring complex multi-step reasoning (e.g., GUI Agents).
Output-oriented activation (CoME): Such as Xiaomi's CoME architecture, which adopts "output-oriented activation," aligning the activation of experts with specific reasoning stages (e.g., screen summarization, sub-task planning, action decision-making).
Scenario feature routing (EMoE): BYD's autonomous driving solution adopts EMoE, where its Router introduces an independent scenario router, directly inputting scenario features (e.g., turning left, turning right, going straight, etc.) and maintaining consistent routing choices across all layers to accelerate model convergence.
Task/modality-aware routing: In Geely's solution, the MoE expert scheduling unit allocates weights through a gating network based on the current task type (e.g., "memory storage," "memory retrieval") and the modality weights of features, and adopts a Top-1 routing strategy to select the dominant expert.
03. Pruning & Quantization
Pruning and quantization are common methods in model deployment to address computing power shortages and increased latency (especially in edge-side model deployment).
Taking Li Auto and XPeng as examples, the focus of pruning and quantization technologies used in different scenarios is as follows:
Pruning and Quantization Technologies in Different Scenarios (Examples: XPeng, Li Auto)
Data source: Public information; Compiled by: ZoZo Auto Research
The Potential of NAR Begins to Emerge
Currently, autoregressive architecture models (AR) are widely used in scenarios such as voice dialogue, autonomous driving planning, and cockpit Agents (e.g., MoonCast, Transfuser, and numerous AR-based world models) due to their mature ecosystem and explicit sequence supervision capabilities. Meanwhile, AR models generally suffer from high inference latency and stability risks (e.g., word skipping, long-segment incomprehensibility, exposure bias, etc.) because they generate tokens/frames sequentially.
In contrast, non-autoregressive models (NAR) have demonstrated a series of advantages in recent years in scenarios such as monolingual TTS, voice dialogue, and driving planning—namely, adapted stability, efficiency, and cost-effectiveness (represented by Xiaomi's ZipVoice-Dialog and OmniVoice, Changan's FlowR2A, Li Auto's TrajHF, etc.). Although the engineering implementation in complex tasks is still limited by basic support such as data and deployment paradigms, resulting in relatively few application scenarios, NAR achieves a speed and quality reversal over AR baselines on multiple benchmarks through parallel generation or flow matching/diffusion paradigms, still counting as one of the promising technical routes.
Comparison of Characteristics Between NAR and AR Models
Image source: ZoZo Auto Research
The generation speed, adapted stability, and computing power-friendly advantages of NAR models can support them as a new paradigm for AI generation and decision-making. Currently limited by the difficulty of native adaptation, data support, and engineering path dependence, they are expected to coexist with AR models in the long-term technical evolution.
Some Technologies Using NAR
Data source: Public information; Compiled by: ZoZo Auto Research
Multi-Agent Technology: Unified Scheduling Core + Different Information Sharing Mechanisms
From 2026 to 2027, model vendors will continue to enhance the capabilities of foundation large models, while hardware vendors are reducing token costs at a rate of 60-70% annually. How to make good use of large model technologies and solve existing characteristics remains the top priority. Reflected in the vehicle side, one of them is the "Always-on Agent." Its mainstream construction method still belongs to multi-agent technology.
In 2026, multi-agent technology tends to converge in technical paths and expands horizontally in application scenarios: it not only adopts the mainstream paradigm of "hub + experts" but also extends from single cockpit and intelligent driving scenarios to cross-system and cross-domain collaborative scenarios (e.g., operations and maintenance, marketing, manufacturing, etc.).
Application Characteristics of Multi-Agent Technology in 2026
Image source: ZoZo Auto Research
Among them, multi-agent technologies from different vendors use different Agent collaboration mechanisms. Overall, the usage frequency of centralized star scheduling types is higher, but using other architectures yields better results in specific scenarios.
Multi-Agent Technologies of Different Vendors in 2026
Data source: Public information; Compiled by: ZoZo Auto Research
Multi-agent systems in 2026:
Do not set up agents with overlapping capabilities, but divide them into multiple types of dedicated sub-agents, including functional, role-based, and cognitive types (including planning, searching, executing, defending, analyzing, learning agents, etc.), each performing its own duties to complete corresponding sub-tasks;
Agents rely on different collaboration mechanisms (standardized communication and state-sharing paths such as A2A protocols, global state sharing, data interaction interfaces, inter-core instruction frames, Memory read/write, streaming caches, etc.) to ensure efficient information flow and consistent state maintenance.
After each sub-agent completes its task, the central module aggregates all sub-task results, integrates and outputs the optimal solution, and feeds it back for execution, achieving parallel processing of tasks to improve real-time performance and relying on division of labor and backup to enhance safety robustness.
Architectural Characteristics of Multi-Agent Technology in 2026
Image source: ZoZo Auto Research
Taking FAW's agent-based engine ECU as an example:
The architecture of this solution divides three physically isolated computing domains on a heterogeneous multi-core SoC, deploying three types of agents: "reaction, planning, and learning":
The reaction agent running on the ASIL-D lockstep core collects sensor signals at a cycle of ≤1ms, processes emergency conditions such as engine knock at the millisecond level, and possesses emergency takeover authority;
The planning agent running on the ASIL-B high-performance core calls the local shared digital twin engine (physical mechanism model + neural network) to simulate multiple sets of candidate control strategies, generates optimal control parameters through multi-objective optimization, and issues them via structured data frames with confidence, validity period, and CRC verification;
The learning agent running on the QM co-processor collects control deviation data in the background and only performs incremental updates on the digital twin model.
The three communicate securely through MPU-protected shared memory partitions and inter-core interrupts, supported by mechanisms such as fault degradation and model verification updates, thereby achieving full closed-loop self-optimizing control of "real-time emergency response—forward-looking planning—online correction" under the premise of ensuring ISO 26262 functional safety.
Architecture of FAW's Agent-Based Engine ECU Solution
Image source: FAW
Workflow of FAW's Agent-Based Engine ECU
Image source: ZoZo Auto Research
In addition to multi-agent technology, proactive service scenarios are also accelerating their implementation with the help of "Claw"-like technologies.
For example, Li Auto's StreamingClaw can be considered an embodied Agent:
Using incremental streaming reasoning combined with master-slave agents, the vehicle no longer processes videos slowly as offline files but updates them while watching, just like a human.
StreamingReasoning is responsible for real-time perception and planning, StreamingMemory uses hierarchical memory evolution to store multimodal experiences, and StreamingProactivity continuously monitors risks and proactively triggers actions.
It realizes monitoring whether the car owner is yawning or playing with a mobile phone while driving and proactively issuing warnings, proactively greeting when picking up the car, and can also identify items in the passengers' hands in real time.
Zooming out to the whole vehicle, the Agent competition has upgraded from a "single product feature battle" to a "full-scenario ecosystem experience battle." Agent services are transforming from "a bunch of features on a configuration sheet" into "a partner that will proactively reach out to you but leave the driving authority to you."
「Contact Information」Mobile Number Same as WeChat ID
Industry Research Department | Mr. Fu 15810027571
Mr. Zhao 18702148304
Data Service Department | Ms. Zhang 13716037793
Strategic Consulting Department | Ms. Han 15810133447
Promotion and Communication Department | Ms. Liao 13718845418
Mr. Du 13910162318