Liu Qingfeng, who has endured three years of grueling hardship, the "lone warrior" is waiting to stage a comeback.

In the history of business, there are certain moments that are destined to be repeatedly reviewed and examined.

In 2000, iFlytek, which had just been established for one year, held its famous "Bantang Conference" in Bantang, Chaohu, Hefei.

At that time, the company's first voice product "Chengyan 2000" encountered a cold market reception, and the company was caught in a dilemma internally. It was the founder Liu Qingfeng who slammed the table and said the sentence that was later written into the company's history: "If you don't believe in the future of speech AI, please leave!"

Twenty-six years later, a similar scenario occurred again at this company. At the upcoming iFlytek 1024 Developer Festival, the company will submit a long-overdue answer sheet: a brand-new flagship general large model based on the new generation of domestic computing power chip Ascend 950.

In the past three years, Spark is the only player among mainstream large models that has been trained entirely on domestic chips throughout the whole process. Restricted by the video memory and bandwidth of the previous Huawei Ascend 910B, its efficiency in training ultra-long context and complex codes is about 5 times lower than that of NVIDIA H200.

This year's market hotspots are precisely focused on code and agents, which means that iFlytek was choked by infrastructure constraints at the moment when it most needed explosive growth.

The emergence of the new-generation Ascend 950 has changed the game chips. Its video memory and bandwidth have increased significantly, and its theoretical efficiency benchmarked against H200 has risen to 65% to 85%. On the premise that computing power is no longer a hard constraint, Spark finally has the underlying conditions to produce top-tier models. To quote Liu Qingfeng's statement at the investor Q&A session, this is a feeling of "sweetness after all hardships".

These four words point to the improvement of training conditions, but for investors, it must point to commercial returns. In fact, as early as two months ago, Liu Qingfeng's wording at the performance briefing was much more heavy.

On August 21, facing inquiries from investors, he frankly admitted: "In this wave of market hotspots dominated by code and agents, iFlytek seems to be somewhat absent." He called this the "regret and pressure" brought to investors.

The pressure has a financial indicator. In the first half of 2026, iFlytek's revenue reached 11.623 billion yuan, a year-on-year increase of 6.52%; the net profit attributable to owners of the parent company was a loss of 204 million yuan. The revenue is still growing, but the growth rate has not met the outside world's expectations for this "AI national team".

From August 21 to October 24, there is a 64-day interval. What iFlytek needs to make up for is the sense of presence that has been missing for most of the past year. Although iFlytek has been continuously launching products closely following the market in terms of code, desktop office and personal agents, it has rarely become the protagonist of discussions.

A routine developer festival has thus been pushed by the management's commitments to become a fulfillment node.

The new model needs to answer technical questions as well as commercial questions: why has a company that has been engaged in artificial intelligence for more than 20 years not appeared at the center of the latest round of AI boom? When will the route it adheres to finally bring returns?

I. The Year of "Absence"

There is a common joke in the industry that if you don't follow AI for a week, you will fall behind, and if you don't follow AI for half a year, everyone will be on the same level.

The speed of technological evolution is extremely fast, and the AI hotspots in 2026 have changed several rounds.

In March, OpenClaw set off the "lobster raising" craze; in April, DeepSeek sounded the counterattack horn of domestic models, accelerating the adaptation of domestic computing power; from June to July, desktop office Agents began to compete for work entrances; by September, Muse, Cue, Dots appeared one after another, and Tongyi released its Personal Agent plan.

In every round of hotspots, new protagonists have stepped onto the center of the stage. As the "AI national team", iFlytek's name rarely appears in prominent positions.

To be fair, iFlytek has participated in every round of technology boom. Training based on domestic computing power has always been the traditional business of the "national team"; in terms of products, on March 12, it launched AstronClaw, and 5 days later launched the desktop assistant Loomy, supporting access to office tools such as Feishu and DingTalk; on September 1, Yuanxing Spark open-sourced the X2.5 series of end-side models.

In terms of release rhythm, iFlytek's follow-up speed is not slow, even earlier than many players who later occupy the center of discussion, but the market has not given the expected popularity.

Taking the AI office entrance as an example, the second-quarter report of Analysys shows that in June, the total visits of 17 mainstream desktop office agents exceeded 60 million times. Among them, Tencent WorkBuddy accounted for 20.97 million times, the domestic version of ByteDance TRAE IDE was 12.79 million times, and Alibaba QoderWork was 7.88 million times. The three Internet giants divided the cake of office efficiency improvement demand.

By September, the discussion of personal agents ignited by Meta still focused on super apps, social chains and ecological distribution. It is not difficult to see that no matter what user group the Agent is for, the key to victory is highly related to distribution channels and existing ecosystems.

Tencent has WeChat and WeCom, ByteDance has Douyin and Feishu, and Alibaba has DingTalk and the Taobao & Tmall consumer ecosystem. These giants do not need to re-educate users, they can directly graft Agents onto the super apps that are already running at high frequency. New tools naturally penetrate into users' daily lives along the original relationship chains and work flows.

In contrast, iFlytek's iFlyRec, which is based on its 20-year voice gene, appears somewhat weak in the office ecosystem.

From the perspective of product form, iFlytek has extremely solid underlying assets. From the iFlyRec APP on mobile phones, the transcription platform on the Web, to the iFlyRec intelligent conference system and voice recorder on the hardware side, iFlyRec has almost achieved the ultimate in single-point experience and engineering implementation.

The prosperity of tools provides depth, but it is difficult to expand boundaries.

The complete link of office is "communication - meeting - resolution formation - task assignment - document collaboration - approval circulation". iFlyRec only cuts into the single node of "meeting", and the work flow stops abruptly here. In order to complete subsequent task distribution and progress tracking, users have to jump out of iFlyRec and return to WeChat, DingTalk or Feishu.

iFlyRec can adapt to the giants' ecosystems, but it will also be marginalized in the context entrance battle when the giants launch corresponding functions - as early as the second half of last year, DingTalk launched Dingtalk A1, and earlier this year, Feishu cooperated with Anker to launch the Voice Recorder Bean.

Based on this, it may be difficult to expect iFlytek to adapt to the broader C-end market. However, in its traditional business of domestic computing power training, iFlytek still let Meituan take the lead - Meituan opened the test of the trillion-parameter LongCat-2.0 on the day DeepSeek V4 was released, with the whole process of training and reasoning relying on domestic computing power, using 50,000 to 60,000 domestic chips.

Since the launch of the large model campaign in 2023, full domestic computing power has always been the most proud label of Spark Large Model, and Liu Qingfeng has repeatedly described the engineering persistence on this road as "lonely brave" in public.

iFlytek has invested a lot of engineering forces to supplement operators, adjust communication, and optimize the training framework. For a long time, the objective constraints of domestic computing power have become the only shield for the company to explain the gap in model capabilities to the outside world.

When "full domestic computing power training" evolved from iFlytek's exclusive position to the industry's default configuration stacked by large factories with massive resources, the uniqueness of the pioneer was inevitably diluted.

As the earliest "technical pioneer" who took the lead in enduring hardships and fully mastered the domestic computing power training process, iFlytek bore the most difficult underlying adaptation and engineering pit-filling costs in the initial stage, and effectively educated and promoted the development of the domestic computing power ecosystem.

After the technical route was verified to be feasible, latecomers with large capital and Internet giants entered the market with far larger computing power scales. Those who worked hard to build the road lost their voice on this road instead.

II. The Ledger of the Absentee

The question is, why does an enterprise with a long history in the AI field, which was the first to bet on domestic computing power and has repeatedly launched new products, always give up the narrative right and traffic high ground to latecomers?

The answer is hidden in the business model that iFlytek has relied on for survival in the past two decades. The business genes that once nourished its growth have shaped its organizational inertia, and finally constructed the company's shackles in the entrance war.

iFlytek has never been a typical C-end Internet company. In the past two decades, the company's resource allocation, business processes and assessment systems have been highly centered on large project systems for government and enterprises (G-end and B-end).

The project system emphasizes customized delivery, long-cycle acceptance and deep scenario binding; while the C-end and desktop entrance war requires extremely low customer acquisition cost, extremely high network effect and the ultimate product iteration speed.

Clayton Christensen proposed in "The Innovator's Dilemma" to use the Resources, Processes and Values (RPV) framework as an analysis tool for enterprise organizational capabilities. Among them, processes and values are born to solve existing problems. Once the environment changes, they will change from "capabilities" to "limitations".

This provides a reference for understanding iFlytek's dilemma. iFlytek has first-class technical resources in China, but its processes and values are deeply customized for large government and enterprise projects.

This kind of organizational inertia is reminiscent of IBM in the era of the rise of cloud computing. The huge and solid "mainframe" customer structure brought the company extremely stable cash flow, but also completely locked its strategic space for rapid transformation to public cloud.

iFlytek can get real money from B-end and G-end contracts, but it is difficult to develop the product muscles to attract C-end users in the public traffic field.

This genetic trace is clearly visible in the financial structure. The financial report for the first half of 2026 shows that iFlytek achieved operating revenue of 11.623 billion yuan, a year-on-year increase of 6.52%.

According to the contraction and transformation strategy of "strengthening the C-end, deepening the B-end, and selecting high-quality G-end businesses", iFlytek's traditional project-based business began to adjust. In the first half of 2026, the revenue of G-end business decreased by 2.65% year-on-year; the combined revenue of B-end and C-end increased by 10%, accounting for 76% of the total revenue.

The improvement of revenue structure has always been difficult. On the profit side, the net profit after deducting non-recurring gains and losses still lost 637 million yuan, while the R&D investment was as high as 3.007 billion yuan, and the R&D expense ratio exceeded 25%. This is a typical financial structure where R&D and delivery overlap, and profits are swallowed up by the process.

In the main battlefield of B-end and G-end, iFlytek still shows extremely strong penetration.

The education track has deeply covered 36,000 schools, and the "Intelligent Medical Assistant" in the medical track has extended to 77,000 grassroots institutions. The monitoring report of Intelligent Hyperparameters shows that in 2024, 2025 and the first half of 2026, the number and amount of winning bids of iFlytek's large model related projects ranked first.

Even on the consumer side, iFlytek's relatively better playing method is still the hardware delivery of "software and hardware integration".

Third-party data shows that its AI learning machine has won 49% of the share in the high-end online market above 6000 yuan, and its sales volume rebounded strongly by 30% in July; AI glasses and intelligent cockpit are also continuously promoted. Hardware has a clear physical unit price and visible gross profit, which is still the extension of the B/G-end logic of "selling equipment and licenses" on the consumer side.

The long-formed technology and organizational genes have constructed the company's competitiveness that is more vertical than horizontal.

After the era of large models arrived, iFlytek sensed the opportunity to break through the boundary. In the first half of 2026, the revenue of the company's large model API and MaaS platform services increased by about 70% year-on-year, and the number of large model developers reached 3.219 million, a year-on-year increase of 111%.

In June, Hefei Yuanxing Spark Co., Ltd. was established independently to undertake large model R&D and commercialization, and plans to raise financing externally. This almost copies the standard playing method of large model unicorns: the subsidiary isolates risks, independently raises financing to obtain high valuation, and provides a private plot for faster innovation.

In this tear between technology and inertia, iFlytek urgently needs to prove that it can not only fight tough battles in the narrow industry depth, but also come up with truly disruptive products.

Under heavy pressure, all eyes are focused on this press conference in Hefei.

III. 1024, A Repricing

The core of iFlytek's "handing in the answer sheet" to the market is undoubtedly the new flagship large model based on the Ascend 950 platform. In addition, there are the progress of independent financing of Yuanxing Spark, as well as the underlying planning of subsequent chips and super nodes.

The management's expectation for Ascend 950 is to reduce the constraints of video memory and bandwidth in ultra-long context training. With chip updates, plus super nodes and training system optimization, there is a chance to catch up with the previously suppressed code and agent capabilities.

Throughout the history of technology and business, old technology giants facing the impact of epoch-level technology waves often go to two completely different fates.

One is the IBM template. When the cloud computing and mobile Internet waves hit, although IBM had deep technical accumulation and a huge enterprise customer base, it eventually completely fell behind in the public cloud market and became a slow company providing traditional IT integration services due to slow strategic actions and huge organizational inertia.

The other is the Microsoft template. After missing the mobile Internet era, Microsoft's Satya Nadella proposed the "mobile first, cloud first" strategy, reorganized the organizational structure and business focus, tilted the company's resources to the Azure cloud business, and bet on OpenAI keenly.

The company finally achieved an epic comeback in the history of business and returned to the peak of global market value.

To judge which path iFlytek will take after October 24 this year, the core depends on two indicators: first, whether the new flagship model based on Ascend 950 can truly enter the first echelon at home and abroad in terms of hard indicators; second, based on this model, whether iFlytek can rebuild the developer ecosystem and the narrative confidence of C-end users.

The problem is that even if both indicators are met, iFlytek's position will not change within a single press conference. Referring to Microsoft, Nadella took office in 2014, and Azure took four full years to establish its industry position, and the bet on OpenAI was not proven until after the end of 2022.

Perhaps the real significance of this developer festival for iFlytek is not a "turnover battle" that achieves success in one action, but to provide the market with an opportunity to re-examine and reprice it.

If the capabilities of the new model fail to meet market expectations, the "sweetness after all hardships" that Liu Qingfeng looks forward to will probably have to wait for the redemption of the next chip.

This article is from the WeChat official account "Yinxing Finance" (ID: yinxingcj),