iFlytek releases end-cloud integrated Spark X2.6; can its voice legacy support a new business in the physical world?

On October 22, at the main forum of the 2026 iFLYTEK Global 1024 Developer Festival, iFLYTEK released the device-cloud integrated Spark X2.6.

How the model enters cars, glasses, and robots has become the next question this company, which started with speech technology, must answer. This path did not begin today. iFLYTEK has long provided intelligent interaction products to automakers and also operates a robot super-brain platform. Now, it hopes to combine speech, vision, and large-model capabilities to take on more interaction tasks on more terminals.

But commercial questions follow: which capabilities will customers pay for, and can this revenue improve iFLYTEK's profitability?

01. Why the strategic shift at this moment?

According to the introduction, iFLYTEK Spark X2.6 completed a capability upgrade based on fully domestic computing power, with a model size of 293-A30B, support for a million-token context, and the launch of a domestic ten-thousand-card inference cluster.

iFLYTEK said it was the first to break through the long-horizon task training efficiency bottleneck of domestic chips. Its long-horizon task dynamic parallel training technology, tailored to domestic platforms, solves the problem of low mixed-training efficiency for long-horizon tasks caused by limited memory and bandwidth on domestic chips.

The Fourth Wave noted that at the launch event, iFLYTEK's narrative focus shifted overall toward "bringing models into the physical world."

Its latest financial report may explain iFLYTEK's strategic shift at this moment. In the first three quarters of 2026, iFLYTEK's revenue was 18.187 billion yuan, up 7.05% year on year; net profit attributable to shareholders was 21.8872 million yuan, turning from loss to profit year on year; net profit after deducting non-recurring gains and losses was a loss of 763 million yuan, with the loss expanding 125.79% year on year; the return to positive attributable net profit relied on 785 million yuan in non-recurring gains and losses, including 471 million yuan in government subsidies. The book turnaround still depends on non-recurring gains and losses, and the profitability of day-to-day operations needs continued improvement.

Investment is also increasing. The company said in its third-quarter report that R&D investment in the third quarter exceeded 1.7 billion yuan, an increase of nearly 400 million yuan year on year, which was the main reason for the decline in net profit after deducting non-recurring gains and losses that quarter. At the same time, net cash flow from operating activities in the first three quarters was 264 million yuan, and government subsidies were 2.4 times the deducted non-recurring profit. The collection side is also under pressure: as of the end of 2025, the book value of accounts receivable was 16.296 billion yuan, equivalent to 60% of annual revenue, with bad debt provisions of 4.342 billion yuan already made.

These figures show that for iFLYTEK, what needs to be solved now is how to make continuously invested technology bring more stable revenue and profit. Cars, glasses, and robots provide new application space, but entering these markets does not in itself mean finding a high-margin growth curve.

The entire speech sector faces the same structural problem. Engineering accumulations such as recognition accuracy, noise resistance, and low latency are still real strengths, but large models bypass the extreme demand for "accuracy" through semantic understanding and cover modules that previously required dedicated procurement with general capabilities.

In addition, the pressure on the speech sector is real. Doubao Input Method used Seed-ASR to shift the battlefield from "hearing accurately" to "understanding," with downloads exceeding 50 million in the first week after launch in November 2025 and 12 million daily active users; AISpeech's revenue compound growth rate from 2023 to 2025 was only 12.98%, with gross margin above 60% but losses for consecutive years, and its period expense ratio has long been above 68%; Unisound's 2025 revenue was 1.211 billion yuan, up 29% year on year, with large-model business revenue jumping from 51.87 million yuan to 610 million yuan, but the full-year net loss was still 329 million yuan; SoundHoundAI fell 61% from its 2025 high, and Cerence dropped 56%.

Speech companies used to sell "speech technology itself," but now customers need interactive Agents that "can understand and can get things done," and the latter is the factory setting of large-model companies.

iFLYTEK is also adjusting its revenue structure. In the first half of 2026, iFLYTEK's open platform revenue was 3.705 billion yuan, surpassing the education segment for the first time, and large-model API and MaaS revenue grew about 70% year on year, showing that it is also shifting toward "selling model calls," but the open platform's gross margin is only 20.21%, and the new curve has not yet grown thick enough to replace the old curve.

02. The "second curve" remains to be verified

Entering the physical world offers iFLYTEK another possibility for charging: expanding from providing a certain speech capability to taking responsibility for an entire interaction solution. This requires it to simultaneously handle coordination among models, sensors, device performance, and the actual environment.

Take robots as an example. When several people speak at the same time and there is device noise nearby, it must first determine who is speaking to it and when it should respond, and only then understand the content and execute the task. Interaction in cars and on glasses has similar problems: users will not always issue commands in standard sentence patterns in a quiet environment with a stable network.

A robot product from an ecosystem company under iFLYTEK.

Lip-movement recognition noise reduction on iFLYTEK AI glasses can illustrate the value of this kind of engineering work. According to the company's introduction, the camera captures lip-shape information and combines it with sound collected by microphones to help the system identify the target speaker in noisy environments. Adding visual information beyond sound is intended to give the model more reliable input.

This gives iFLYTEK's years of accumulated acoustics and interaction technology room to be used. However, whether a feature can become a reason for customers to purchase still depends on whether it is stable across different hardware and environments, and whether deployment and maintenance costs are worthwhile. Model companies and hardware manufacturers can also enter this market, and iFLYTEK still needs to prove its advantages in actual delivery.

iFLYTEK's existing terminal business is the foundation of this route. The semi-annual report shows that in the first half of 2026, the smart car business added more than 4.5 million units of pre-installed intelligent products shipped. Cars provide practical delivery experience, and the robot platform has also accumulated a group of partner manufacturers. The company previously said on an investor interaction platform that its robot super-brain platform has served more than 500 intelligent robot manufacturers.

But the number of manufacturers served cannot yet answer how big this business is. Platform trials, technical cooperation, and bulk procurement have different commercial value; how many robots actually carry the platform, how much revenue each device contributes, and whether customers renew are what matter for subsequent revenue. The scale of the existing car business cannot be directly applied to robots.

Device-cloud collaboration is a path iFLYTEK is trying in order to control costs. Putting tasks suitable for local processing on the device side and handing complex tasks to the cloud is expected to reduce some cloud calls and network dependence. But on-device models also consume computing power, memory, and battery. Whether the cost of the whole machine can be reduced needs to be calculated based on specific devices and tasks; the fact that computing power, models, and interaction technology can be reused does not mean that R&D and adaptation costs for various terminals will naturally decline.

More direct competition comes from customers themselves. AgiBot has already released the WITA-Omni model for embodied interaction and in September 2026 demonstrated an application in which a robot identifies interaction targets and coordinates language and actions. Robot manufacturers are developing interaction capabilities in-house, which means iFLYTEK must both serve them and face the possibility that they expand the scope of their own development.

In-house development and external procurement are not necessarily mutually exclusive. Leading manufacturers can develop models themselves while purchasing speech, noise reduction, or specific modules; manufacturers with limited R&D resources may also be more willing to buy complete solutions. How much business third-party suppliers can retain depends on the cost of customers developing it themselves and on the advantages of external solutions in effectiveness, delivery, and maintenance.

This is also where iFLYTEK's "second curve" truly needs to be verified. It already has technology and customer entry points. The next step is to prove whether these cooperations can form sustained orders and whether, after delivery scale expands, revenue can cover model, computing power, and adaptation costs. Only when these figures gradually become clear will moving toward the physical world turn from a technology route at a launch event into a new business in the financial report.

This article comes from the WeChat public account "The Fourth Wave," author: Lu Yang, published by 36Kr with authorization.