The deep integration of AI large models, agents, and vehicles is pushing AI cockpits into the fast lane.
Across the market, a growing number of automakers are rolling out AI cockpits in new models, positioning them as key selling points to sharpen product appeal and elevate the user experience.
Meanwhile, upstream players are accelerating their push. From chip and OS makers to Tier 1 suppliers, cloud providers, large-model developers, and app ecosystems, suppliers across the board are doubling down—each vying for dominance in this intelligent wave.
iFLYTEK, drawing on more than two decades of deep expertise in smart voice and AI, has built a software-hardware integrated AI cockpit tech stack across models, hardware, and harness engineering. Its Spark Smart Cockpit 3.0 went into mass production the moment it was released—making iFLYTEK one of the first in this industry race to close the loop from technology to volume production.
Smart Cockpits Evolve: Enter the AI-Native Era
Looking back over the past decade or so, the evolution of smart cockpits falls roughly into three phases.
The first was the infotainment cockpit, spanning roughly 2010 to 2015.
Cockpits of this era centered on physical buttons and a small central display. Features like navigation, Bluetooth, and music gradually appeared, though all were fairly basic. Crucially, the feature set was largely locked in at the factory—adding new functions post-delivery was nearly impossible.
The second phase—smart connected cockpits—ran from about 2016 to 2023.
Touchscreens began replacing physical buttons, center displays grew larger, and screens multiplied inside the cabin. Voice interaction, vehicle-to-cloud connectivity, and OTA updates hit the road in succession. Cockpits shifted from closed systems to open platforms, loading more mobility and infotainment apps.
iFLYTEK' s Flyfish system, for instance, was supporting multi-zone and multimodal interaction as early as 2018.
By 2020, the Flyfish system supported multiple languages, each covering seven major domains—vehicle controls, music, radio, navigation, commands, Bluetooth calls, and lifestyle services—across more than 200 interaction scenarios. Users could answer calls, play music, or set a route by voice, achieving a basic “speak, don' t touch” experience. iFLYTEK' s pre-install user base surpassed 30 million, partnering with over 90% of China' s mainstream independent and joint-venture automakers.
Still, the underlying interaction logic remained unchanged: users still had to find the right app for the task. Voice acted more like a shortcut—recognizing standard commands but struggling with complex natural language or cross-app tasks.
The third phase—AI cockpits—is now in its early stages of rollout.
The fundamental shift here isn' t just more features; it' s AI evolving from a plug-in capability into the cockpit' s native foundation. Against this backdrop, AI cockpits diverge from their predecessors in five significant ways:
In service delivery, passive chatbots give way to proactive assistance. The system reads the user' s state and driving context, serving needs before the user even asks—rather than waiting for a command.
In processing logic, simple feature stacking yields to task-driven loops. Users no longer need to remember which app to open or toggle between interfaces; they simply state their intent, and the system decomposes the task, allocates resources, and executes.
In perception, single-mode voice or vision shifts to multimodal fusion. The system doesn' t just listen—it combines imagery, occupant state, and cabin environment to gauge intent, even reading emotions. It transitions from a tool to an empathetic companion.
In underlying architecture, siloed domains merge under unified orchestration. Previously, navigation, climate, and media operated independently; now, a central AI agent coordinates them all. A single sentence can mobilize multiple systems in concert.
In runtime, edge-cloud collaboration has become the mainstream architecture. The edge handles fast responses to high-frequency commands, protects privacy, and works offline; the cloud tackles complex reasoning and multimodal interaction. Together, they let a vehicle' s capabilities keep growing after delivery.
Image source: iFLYTEK
iFLYTEK' s newly launched Spark Smart Cockpit 3.0 is a prime example of this new paradigm.
Positioned as the industry' s first full-modal, human-like interactive cockpit, Spark Smart Cockpit 3.0 is built on the SOPHIA architecture. It forms a closed loop across four systems: full-modal perception, cognitive decision-making, multidimensional expression, and memory evolution.
Specifically, full-modal perception lets the car both hear user intent and see cabin conditions. The decision layer' s Master Agent acts as the vehicle' s unified dispatch hub—interpreting intent, planning tasks, allocating capabilities, and overseeing execution. For expression, a tailored “Spark Family” helps automakers customize brand personas. Meanwhile, the memory system continuously accumulates user habits while distinguishing the preferences of different passengers—delivering a truly personalized experience.
For users, Spark Smart Cockpit 3.0 currently delivers three core values: an empathetic companion who understands you, a proactive automotive butler, and an immersive audio-visual master.
As early as 2023, iFLYTEK released Spark large model-based automotive applications and upgraded its smart cockpit solution. As an early AI cockpit prototype, the Spark cockpit achieved freer, more human-like interaction built on multi-turn, multi-person, and multimodal exchanges—making iFLYTEK a trailblazer in securing a foothold in AI cockpit technology.
Long term, the goal of AI cockpits is to evolve the car into a “super-vehicle agent”—one with comprehensive cognition, proactive service, and continuous evolution.
Software-Hardware Integration Is the True Ticket of Admission
Putting large models on wheels has undeniably ushered smart cockpits into the AI-native stage. Yet an AI cockpit is not merely a large model stacked onto a smart cockpit. For these capabilities to truly land, a cohesive system of software-hardware coordination is essential.
Consider the software layer first.
A complete AI cockpit software stack roughly comprises four parts. At the base: an AI development platform and agent orchestration, handling model development, deployment, and agent organization or task allocation. Above that sits Model-as-a-Service (MaaS), managing the collaborative invocation of large language, multimodal, and voice models across cloud, edge, and edge-cloud—solving “where capabilities come from.” Next up are AI applications: smart voice, AI assistants, navigation Q&A, and content generation—features users directly engage with. At the top: AI operations services, supporting continuous updates and maintenance after models and apps go live.
Internet and cloud giants like Volcengine, Alibaba, Baidu, and Tencent are entering the AI cockpit race primarily through software and services—each with a different emphasis. Some excel at foundational platforms and compute; others lead in large-model tech; still others boast rich app ecosystems.
Overall, this layer dictates the cockpit' s intelligence and the richness of its in-cabin applications.
But getting these capabilities into vehicles at scale ultimately depends on hardware—compute chips, cockpit domain controllers, sensors, displays, and acoustic components.
For one, hardware is the physical substrate for software and models.
Edge-side cockpit models range from hundreds of millions to tens of billions of parameters; without matching compute, bandwidth, and memory, they simply cannot run smoothly. Likewise, rich in-cabin apps and ecosystems cannot close the interaction loop without screens and acoustic hardware.
In other words, hardware ultimately caps the AI cockpit experience.
Image source: iFLYTEK
For another, hardware does two things software cannot: capture information from the physical world and translate digital commands into physical action.
However smart a model is, it needs cameras and microphones to capture sights and sounds before it can “see” or “hear.” And when the model issues a command, it takes motors to adjust seats, climate, and windows to manifest that action in the physical world. Information flows in from the physical realm, and commands flow out to it—no amount of code can replace these two endpoints.
In this sense, software-hardware integration isn' t optional for AI cockpits—it' s the price of entry. Separate the two, and neither side can constitute a complete AI cockpit on its own.
Edge-side model parameters must align with chip compute, bandwidth, and memory; perception and execution must land on physical devices. On top of that, automakers must solve for automotive-grade compliance, cost control, and production timing—ensuring software runs stably and efficiently on automotive hardware. Only when both are achieved can AI capabilities truly be redeemed inside the cockpit.
Yet software-hardware integration is merely the threshold, not the finish line.
Cross it, and the industry' s competitive center of gravity shifts toward mass production and ecosystems. The larger a player' s production scale and the more valid data it accumulates, the faster its models will evolve. The richer the app ecosystem, the better the cockpit can retain users.
Ultimately, integration determines whether a tech can get on the road; production and ecosystems determine how far it goes once it' s there.
From Tech to Reality: Mass Production Separates Contenders from Pretenders
Over the past decade, the smart cockpit space has crystallized into a relatively clear market structure: Desay SV leads in cockpit domain controllers, Qualcomm in domain-control chips, iFLYTEK in in-car voice, and Foryou Multimedia in HUDs.
But as the AI cockpit era dawns, the rules are changing. Without hardware, even the most powerful models and ecosystems remain trapped in the digital realm. Without software, the most reliable hardware cannot spontaneously generate intelligence.
That is why, facing the AI cockpit' s new trajectory, iFLYTEK chose to deepen the integration of software and hardware along its strongest chains—voice and large models—building a software-hardware integrated AI cockpit tech stack.
Notably, iFLYTEK is one of the rare full-stack players in the AI cockpit space with in-house R&D across large models, agent orchestration, smart voice, acoustic systems, and edge-side compute hardware.
Image source: iFLYTEK
On the software side, iFLYTEK' s core product is the Spark Auto Large Model, SparkAuto. Unified by a Master Agent core, it maps technically to the SOPHIA architecture' s four systems—full-modal perception, cognitive decision-making, multidimensional expression, and memory evolution—and splits into two series: cloud-side SparkAuto and edge-side SparkAuto-EMM.
The edge-side SparkAuto-EMM spans multiple sizes—2B, 4B, 7B, and 35B—and covers over 100 cockpit scenarios. It can be flexibly configured for different automakers and vehicle tiers, balancing experience, cost, and security through edge-cloud collaboration.
iFLYTEK also holds the voice-base large model SparkAudio, along with industry models for education and healthcare, providing cockpits with richer content and capability sources.
Take healthcare: powered by the Master Agent and iFLYTEK' s Xiao Yi medical assistant, Spark Smart Cockpit 3.0 brings the “measure, ask, diagnose” loop into the daily commute—linking the car, home, smartphone, cloud, and wearables from Huawei and Apple.
On the hardware side, iFLYTEK has laid out edge-side compute hardware and acoustic systems like microphone arrays and speakers around cockpit scenarios—pairing seamlessly with its edge-side models.
Image source: iFLYTEK
In edge compute specifically, iFLYTEK just released two AI Boxes that directly host edge-side large models: the SparkAuto BOX, in Lite and Ultra series. The SABOX-Lite, built on the RK1828 platform, delivers 20 TOPS of compute and runs a 4B model; the SABOX-Ultra packs 200 TOPS and runs a 35B model.
By packaging large models and compute into standardized boxes, iFLYTEK transforms models from “algorithms running in the cloud” into “software-hardware integrated products that ship with the car.”
Beyond that, iFLYTEK designed and developed the Spark Smart Cockpit 3.0 Harness engineering around the SOPHIA architecture—turning SOPHIA from an internal diagram into an industry-grade engineering base that can be stably mass-produced and continuously evolved.
Even the most comprehensive tech stack, however, must be proven in real vehicle programs. From co-defining requirements with automakers to passing automotive-grade validation, and from mass-production delivery to subsequent iteration—every link determines whether this software-hardware capability can truly be installed in a car and run reliably.
This is precisely iFLYTEK' s strength.
Backed by over two decades of deep cultivation in in-car voice, iFLYTEK has amassed mature engineering deployment experience and broad client relationships in automotive-grade validation and mass-production delivery. Installation volume is the most direct proof.
According to Gasgoo Research Institute' s smart cockpit configuration data, iFLYTEK secured the No. 1 spot for in-car voice supplier installations in 2024 with 8.276 million standard-fit units—over 5 million units ahead of the runner-up.
Image source: Gasgoo
In the two years since, iFLYTEK has maintained a commanding lead in the in-car voice market. In the first eight months of this year, its voice installations in China' s passenger-vehicle market reached 4.361 million units, capturing a 42.7% share. For context, runners-up AISpeech and Cerence logged 2.595 million and 1.056 million units over the same period, holding 25.4% and 10.3% of the market, respectively.
Overall, iFLYTEK' s automotive pre-install products have reached 81 million cumulative units, partnering with over 50 automakers across 60-plus countries and regions. It supports 32 core languages and has delivered over 500 production models.
Consider that in-car voice is now the core interaction portal for smart cockpits.
Traditional cockpit interaction centers on physical buttons and touchscreens—users must locate controls and navigate menus to act. But as cockpit functions multiply, app-style layered menus only drive up interaction costs.
Voice interaction, by contrast, naturally fits the driving scenario: hands-free, eyes-free, interruptible, and continuous. It perfectly maps to high-frequency essentials like navigation, vehicle controls, calls, music, weather, charging, and parking.
Especially over the past two years, the rapid evolution of AI large models has upgraded voice from “understanding commands” to “grasping intent”—pushing smart cockpits toward an app-less paradigm. Users simply state what they want done; the system executes directly, eliminating the need to toggle between menus and apps.
Voice has thus evolved from a feature portal to the primary portal dispatching all vehicle tasks. Controlling this gateway means capturing the first touchpoint of user intent.
More importantly, large-scale real-world in-car interaction generates rich scenario data, feeding back into model iteration and making models more adaptable to cabin environments. Meanwhile, extensive cross-brand and cross-model adaptation accumulates the engineering expertise needed to run the same model stably across different hardware and vehicle types.
Together, these elements form iFLYTEK' s first-mover advantage in the AI cockpit race.
Powered by this wealth of production experience and strong product capabilities, Spark Smart Cockpit 3.0 achieved mass production upon release. It has already been deployed in models from Chery, GWM Wey, and Dongfeng Nissan, rapidly securing a key position in the AI cockpit transformation.
Going forward, as smart cockpits fully transition into the AI era, true clout will ultimately be established by installation scale and mass-production delivery. iFLYTEK' s task is to continuously convert its first-mover edge into delivery and experience advantages—maintaining its lead in this long-distance race.
