As AI technology has been densely integrated into vehicles over the past few years, cars have indeed become smarter: they can speak, answer, and handle voice-controlled driving, navigation, and music requests instantly.
But 'being able to converse' does not mean 'being able to get things done.' Once applied to real vehicle-use scenarios, the shortcomings of in-vehicle AI remain clear: insufficient proactivity, fragmented systems, stiff interaction, and closed scenarios.
The root cause lies in how AI is integrated into vehicles.
In the first half, AI was more like features added into cars one by one. This 'bolt-on' intelligence was layered onto the existing whole-vehicle architecture, unable to break through the barriers among the cockpit, intelligent driving, and vehicle control, so the experience was naturally fragmented. To fundamentally solve these pain points, the industry must move toward AI-native development, with AI as the core, driving the evolution of the whole-vehicle electronic/electrical architecture and reconstructing the in-vehicle service ecosystem.
iFLYTEK is doing exactly this.
On October 21, iFLYTEK held a smart car product and technology launch event in Beijing under the theme 'Pursuing Light, Evolving Without Boundaries.' During the event, iFLYTEK not only officially launched the industry's first omnimodal human-like interactive cockpit, SparkAuto Space 3.0, but also presented a complete AI-native vehicle 'intelligent foundation' across five layers: architecture, model, hardware, engineering, and ecosystem.
iFLYTEK also announced that brands including Qirui, Changcheng Weipai, and Dongfeng Nissan have taken the lead in adopting selected experiences of SparkAuto Space 3.0, officially opening a new stage of AI-native development for smart cars.
SparkAuto Space 3.0: Letting AI Grow from the Vehicle's Underlying Architecture
If the first half of AI cars competed on 'deploying quickly and stacking many features,' the second half competes on 'landing solidly and being used well.'
Over the past few years, cockpit screens have grown larger and feature menus have become longer, but users actually use only a few features at high frequency. This shows that the number of features does not equal the quality of the experience.
In iFLYTEK's view, what truly matters in the second half is valuable scenarios and services that are genuinely useful to users.
Applying these roles to the product, SparkAuto Space 3.0 offers three answers: an emotional companion that understands you better, an automotive concierge that gets things done, and an ultra-immersive audio-visual master.
As the industry's first omnimodal human-like interactive cockpit, SparkAuto Space 3.0 is built on the SOPHIA architecture (Human-like Intelligence Omnimodal Smart Car Interaction Architecture). Referencing human perception, thinking, expression, and memory, it establishes four systems: omnimodal perception, thinking and decision-making, multidimensional expression, and memory evolution, letting AI grow directly into the vehicle's underlying architecture.
Start with the 'brain': Master Agent is the unified scheduling hub for the whole vehicle.
In the past, a user's single sentence was often split among multiple isolated features for processing, but real needs have never been single-point.
For example, a sentence such as 'I'm hungry; find a Hunan restaurant with few customers and a high rating, and also book movie tickets for 8 p.m.' contains dining, ticket purchase, taste preference, and time constraints.
In the past, the in-vehicle system was only responsible for 'opening features,' while the rest required users to switch back and forth among several apps, and the system could not automatically complete the transaction loop.
The Master Agent of SparkAuto Space 3.0, as the 'brain' that schedules tasks, coordinates user intent understanding, task planning, task scheduling, and task supervision, so it can automatically decompose and order a single sentence and directly output execution results.
Being able to schedule is not enough; it must also remember, feel, and express.
In memory, SparkAuto Space 3.0 implements more than 68 memory dimensions, supporting seats, rearview mirrors, frequently listened podcasts, and other items, so users return to the previous state as soon as they enter the vehicle. With a phrase like 'that restaurant from last time,' there is no need to provide the address again.
Considering that personalization and data security are often in tension, the system also sets two 'privacy locks': memory queries require owner authentication, and highly sensitive memories are stored only on the device side and not uploaded to the cloud; users can view, modify, and delete them at any time. This is equivalent to first laying a solid compliance foundation before adding personalization on top.
In perception, voice interaction extends from inside the vehicle to 30 meters outside, and special-vehicle sound recognition reaches up to 200 meters. Device-cloud multimodal perception covers more than 100 cockpit scenarios, with multimodal perception accuracy as high as 95% and fuzzy understanding accuracy of 93%. The device-side enhanced sentinel mode alone covers 16 categories of risk events, delivers assessment results for a single event within 4 seconds, and achieves recognition accuracy as high as 90%.
In expression, the 'Spark Family' helps automakers customize brand personas across five dimensions: personality, image, action, voice, and language. For multi-person scenarios, the multi-person omnimodal full-duplex solution supports 'one person wakes it, and the whole car can speak,' and can distinguish 'who is speaking' and 'to whom it is speaking,' without interrupting or cutting in.
A smarter brain, more reliable memory, keener perception, and more human-like expression establish 'an emotional companion that understands you better.' But iFLYTEK wants more than this: it also wants the vehicle to 'get things done' and do things that 'understand you better.'
This is the automotive concierge that gets things done, managing everything from travel and parking to health.
For travel, SparkAuto Space 3.0 deeply integrates travel, vehicle use, and local life services, supporting automatic gate lifting when entering expressways, password-free payment when leaving, route reminders to pick up packages, automatic ordering before arriving at a store, and handling queues, movie tickets, hotels, and scenic spots with a single sentence.
For parking, the newly launched Suixinbo Agent, integrated with the Master Agent, deeply fuses the 'brain' with intelligent driving, enabling autonomous parking, on-demand vehicle moving, and summoning for pickup through 'Suixinbo.'
This also offers another approach to cockpit-driving integration: true integration need not be the merging of hardware domain controllers; it can also be scheduling at the decision-making layer by a unified 'brain.'
For health, relying on the Master Agent and Xunfei Xiaoyi, SparkAuto Space 3.0 brings the 'measurement, inquiry, and diagnosis' loop into daily commuting and connects the vehicle, home, mobile phone, cloud, and Huawei and Apple wearable devices. Behind this is Xunfei Xiaoyi's cumulative 200 million health consultations, more than 35 million users, 2,600 diseases, and knowledge of more than 60,000 medicines.
The 'ultra-immersive audio-visual master' is also the result of AI redefining the experience.
Specifically in SparkAuto Space 3.0, the biggest highlights are the industry's first mass-producible gallium nitride high-power acoustic system and the AI ultra-immersive surround algorithm, which can reconstruct ordinary stereo into 7.1.4 panoramic sound in real time, while reducing the weight of the entire system by 30% and adapting to more than 30 sound effects.
The newly launched smart cockpit quietness system, relying on the industry's first ultra-low-latency amplifier hardware and a new noise-cancellation algorithm, enables AI to automatically complete the entire process of 'noise capture → real-time analysis → reverse cancellation,' reducing overall vehicle noise by 9 dB(A). The cockpit becomes nearly as quiet as a residential community late at night, allowing mainstream-priced models to enjoy luxury-car-level quietness.
In addition, iFLYTEK also launched the Smart Sound Field Zoning System 3.0 and the Super Acoustic Headrest System, moving cockpit acoustics from a hardware arms race of 'stacking speakers' into a new stage where AI redefines the sound field.
Overall, from an emotional companion and automotive concierge to an audio-visual master, SparkAuto Space 3.0 presents not a single feature becoming 'smarter,' but intelligence grown from the underlying architecture. It is this 'native' nature that transforms the vehicle from a tool that passively executes commands into a whole-vehicle intelligent agent that truly has emotions, gets things done, and provides companionship.
Device-Cloud Integration: Using One 'Automotive Brain' to Connect Full-Stack Capabilities
For AI-native smart cars, embedding AI at the underlying architecture is only the first step; what truly determines the upper limit of the experience is a large model born specifically for automotive scenarios.
Over the past two years, general-purpose large models have become sufficiently usable in smartphone and office scenarios, but a car is not a smartphone: the cockpit, intelligent driving, and vehicle control experiences must be continuous, user memory must be uninterrupted, responses must be fast enough, the cost of high-frequency calls must be controllable, and operations involving vehicle control must be absolutely safe. None of these can be solved by simply putting a general-purpose cloud large model into a vehicle.
Insight into this, iFLYTEK formed a full-link AI breakthrough team to advance the R&D of automotive-specific large models. At this event, the SparkAuto automotive large model was officially released.
The SparkAuto automotive large model uses Master Agent as its unified core. Technically, it corresponds to the four systems of the SOPHIA architecture: omnimodal perception, thinking and decision-making, multidimensional expression, and memory evolution. It is divided into two series: cloud SparkAuto and device-side SparkAuto-EMM.
Among them, cloud SparkAuto 1.0 adopts Cabin-RL cockpit sandbox simulation and safety-aligned reinforcement learning, with four capabilities: unified scheduling of atomic capabilities, continuous memory, complex task processing, and safety and trustworthiness in automotive scenarios.
It performs low-latency inference for a 350K ultra-long context. Based on H20 card measurements, the cache hit rate exceeds 99.5%, first response is less than 800 milliseconds, and computation speed exceeds 200 tokens/s. In terms of experience, ultra-long conversations do not drop out, and users hardly need to wait before speaking.
Device-side SparkAuto-EMM covers multiple sizes of 2B/4B/7B and more than 100 cockpit scenarios.
At this event, iFLYTEK further debuted the 35B-A3B ultimate device-side multimodal large model. From the initial design, this model aligns with the cloud; when connected to the network, it can also call cloud sources, but core inference computation is still completed on the device side, combining advantages such as offline availability, ultra-fast response, privacy security, and token freedom.
On NVIDIA Jetson Thor T5000, 35B-A3B measured first-frame latency below 650 milliseconds, decoding speed above 80 tokens/s, and peak memory reduced by 75%.
Why does iFLYTEK pursue device-cloud integration? Because at the current stage, betting on only one side is difficult to make work.
Betting entirely on the cloud cannot avoid token cost, response latency, network disconnection, and privacy; betting entirely on the device side has capability limits. The value of device-cloud integration lies precisely in finding a mass-producible balance among experience, cost, and security.
iFLYTEK co-founder and president Wu Xiaoru provided a division-of-labor breakdown: under device-cloud collaboration, about 60% of problems can be completed independently on the device side, 30% are solved through device-cloud collaboration, and only the remaining about 10% of complex tasks go to the pure cloud. 'There is an important prerequisite here: the device-side performance must be good.'
This means that in the second half of large models entering vehicles, device-side capability is the true 'winning move.'
But no matter how strong the device-side model is, it must be installed in the vehicle through hardware. At this event, iFLYTEK simultaneously launched two AI Boxes that directly host device-side large models—SparkAuto BOX, including the Lite and Ultra series.
Among them, SABOX-Lite is built on the RK1828 platform, offers 20 TOPS of computing power, supports multi-chip cascading and 4B large models, and has been deeply trained for vehicle scenarios. SABOX-Ultra offers 200 TOPS of computing power, carries a 35B large model, has been deeply trained for vehicle scenarios, and supports an ultimate device-side interaction experience.
In the past, large models entering vehicles were long stuck in the middle ground: algorithms were in the cloud, but deployment required customization for each vehicle model. Models were not lacking; what was lacking was a carrier that could be replicated at scale.
What SparkAuto BOX tries to fill is exactly this gap: putting the large model and computing power together into a standardized box, transforming the model from 'an algorithm running in the cloud' into 'a software-hardware integrated product that can ship with the vehicle,' so automakers do not need to re-adapt for every model.
With both the model and hardware ready, there is still a hurdle between 'being able to run' and 'daring to mass-produce.'
Large models are essentially probabilistic models, not deterministic programs. If a smartphone answers a sentence incorrectly, users can laugh it off. But once used in a fast-moving vehicle, any hallucination or misjudgment could lead to a safety accident. No matter how strong a model's capability is, it cannot be directly equated with reliability.
Harness engineering is precisely meant to bridge this gap.
Centered on the SOPHIA architecture, iFLYTEK designed and developed the SparkAuto Space 3.0 Harness engineering, turning the SOPHIA architecture from an internal architecture diagram into an industry-level engineering foundation that can be stably mass-produced and continuously evolved.
This engineering clearly delineates responsibilities between the vehicle side and the cloud: the cloud Master Agent is responsible for 'thinking,' handling task planning, intent understanding, multimodal Q&A, and reflection optimization; the vehicle-side EdgeAgent is responsible for 'implementation,' handling event perception, proactive services, offline fallback, and result arbitration. The two collaborate through intelligent routing and state synchronization, and are safeguarded by security protection, context engineering, and layered memory.
Put more plainly, the model thinks in the front, while the system backs it up at the rear, ensuring that every output is checked.
However, models, hardware, and engineering answer whether AI can be built into a vehicle; truly large-scale deployment on the vehicle side still depends on the ecosystem.
Open Co-Creation: From a 'China Foundation' to a 'Global Foundation'
For the second half of AI cars, iFLYTEK has built a complete industry-level intelligent foundation.
But to truly move smart cars toward a high-intelligence, high-value AI-native era, relying solely on iFLYTEK is not enough.
Liu Junfeng of iFLYTEK put it directly: in an era where large models and agents flourish, relying on a 'full-suite' model or going it alone does not work; deeper open co-creation is necessary.
The logic behind this is clear: in the traditional model, once a supplier delivers products to automakers, cooperation basically ends, but AI-native vehicle capabilities require continuous updates, so this relationship is no longer sufficient.
Insight into this, iFLYTEK proposed full-link co-creation and delivery centered on Master Agent: from consensus on definitions and co-creation of configurations to collaboratively complete product design matching a vehicle, to joint validation and mass-production operations to jointly define vehicle acceptance and evaluation methods, and further to full-lifecycle technology iteration and updates, transforming the previously fragmented supply-demand relationship into a long-term co-building chain.
Considering that as AI-native vehicles carry more and more applications, travel, life, entertainment, content, and health cannot all be handled by one company, iFLYTEK also launched the 'Linlang' SparkAuto automotive ecosystem agent platform, allowing partners to connect through standard protocols, combining Master Agent capabilities for different scenarios, customizable for automakers, and subscribable for users.
Currently, the platform has connected ecosystem partners including Zhifubao, QQ Music, Gaode Navigation, and Aiqiyi, aggregated more than 60 ecosystem sources, implemented more than 600 vehicle-onboard projects, and recorded an average daily call volume of 220 million times.
The cooperation between SparkAuto Space and the Zhifubao vehicle assistant is highly representative. In the 'Xunbao Program' jointly created by both parties, SparkAuto Space is responsible for interaction, cognition, and evolution, 'understanding what users really want,' while the Zhifubao vehicle assistant is responsible for 'getting things done and sorting out payments,' with each side managing one segment.
Cooperation with Rockchip Electronics is another approach: domestic chips carrying domestic large models have already mass-produced more than ten vehicle models.
At this event, iFLYTEK further launched the 'Linlang Huicui' program, aiming to add more than 100 excellent agent ecosystem partners and expand more than 100 ecosystem capabilities within the next year.
It is not hard to predict that as SparkAuto Space connects more services, content, and payments, the value of the cockpit will no longer come only from one-time pre-installation licensing, but will extend into continuous subscriptions and service operations. For OEMs, this means long-term revenue from software and services; for ecosystem partners, it is an in-vehicle entry point directly reaching a massive base of car owners.
iFLYTEK not only wants to be the 'all-round foundation' for Chinese cars, but also wants to be the intelligent foundation for global cars.
Currently, Chinese car overseas expansion is moving from complete vehicle export to industrial chain export. Among them, cockpit overseas expansion faces language, culture, service, and compliance hurdles, each more difficult than the last. This is precisely the core capability that iFLYTEK has built over more than 20 years in voice and cognition.
Currently, iFLYTEK's automotive business covers more than 60 countries and regions worldwide, supports 32 core languages, has delivered more than 500 mass-produced vehicle models, has cumulatively provided pre-installed products for more than 81 million vehicles, and has cooperated with more than 50 automakers.
Behind this, for overseas markets, iFLYTEK proposed five product principles: clear hearing, cultural understanding, native expression, getting things done, and maintaining propriety. It has deployed local teams across six global regions, systematically built localized product capabilities, established local security and compliance systems and routine O&M services, and ensured product deployment, data compliance, content security, and long-term service.
Next, SparkAuto Space 3.0 will also comprehensively empower overseas markets, working with Chinese automakers to enter the overseas cockpit large model era and becoming a trusted cockpit foundation co-creation partner for global automakers.
Conclusion
The endpoint of AI entering vehicles is not to make cars better at chatting, but to make them truly able to think, get things done, and grow.
From SparkAuto Space 3.0 to the SparkAuto large model, SparkAuto BOX, and Harness engineering, and further to the Linlang ecosystem and global deployment, iFLYTEK has provided not only a solution, but also a leap path for AI Cars toward 'unified services, shared capabilities, and continuous evolution.'
Of course, 'launching and mass-producing at the same time' is only a beginning.
From scale delivery, to polishing the experience of cross-application services and defining responsibilities, to global compliance operations, the competition in the second half of AI-native vehicles has only just begun.