At the annual 1024 Developer Festival (1024开发者节) in Hefei, one stage unveiled more AI products in an afternoon than many startups ship in a decade. Translator pens, AI glasses, medical diagnostic assistants and a robot-brain platform — all from a company whose founding ambition was simply to make machines speak Chinese well. The question now hanging over iFlytek (科大讯飞) is not whether it can build AI, but whether stretching across so many shelves leaves anything genuinely defensible.
Key takeaways
- iFlytek reported 2025 revenue of 27.1 billion RMB (about $3.8 billion / HK$29.5 billion), up 16% year on year, with net profit of 839 million RMB (about $117 million / HK$912 million).
- The "platform + track" (平台+赛道) strategy spans eight business lines, from AI learning machines (AI学习机) to smart healthcare and autos.
- The Spark (星火) large model is marketed as the only mainstream Chinese LLM trained entirely on domestic compute — a sovereignty story with export appeal.
- Profit quality is thin: government subsidies covered roughly three-quarters of 2025 net profit, and R&D of about 5.4 billion RMB (about $745 million / HK$5.8 billion) is spread across many fronts.
- Global buyers should watch iFlytek less as a model lab and more as a vertically integrated AI-product supplier.
From narrow expertise to a wide shelf
iFlytek was, for its first fifteen years, the textbook "narrow but deep" firm. It won the international speech-synthesis competition fourteen times running and built a moat in speech recognition that few Chinese peers could match. In 2017 the company pivoted to a "platform + track" (平台+赛道) model, betting that speech alone could not be monetized directly and had to be embedded in real scenes.
That decision never really stopped. The company now lists eight business categories: smart education, smart healthcare, the iFlytek Open Platform (讯飞开放平台), consumer hardware, smart cities, telecom operators, smart autos and enterprise AI. Each new tech wave — online education, smart cities, large models, embodied intelligence — pulled iFlytek onto another shelf.
The numbers behind the expansion
The commercial machine is large. Full-year 2025 revenue reached 27.1 billion RMB (about $3.8 billion / HK$29.5 billion), up 16.1% from 23.3 billion RMB (about $3.2 billion / HK$25.3 billion) in 2024. Net profit attributable to shareholders was 839 million RMB (about $117 million / HK$912 million), up 49.9%. The iFlytek Open Platform (讯飞开放平台) reported over 8 million developer teams and more than 800 AI capabilities on offer.
Education is the flagship vertical: the AI learning machine (AI学习机) is reported to hold roughly 49% of China's above-6,000-RMB online segment, a claim attributed to local market data. But scale hides a structural soft spot. Full-year 2024 R&D was about 5.4 billion RMB (about $745 million / HK$5.8 billion), and that sum is divided across more than a dozen product lines.
The Spark sovereignty story
The Spark (星火) model's clearest global hook is autonomy. iFlytek markets it as the only mainstream Chinese general model trained entirely on a domestic compute stack, with algorithms, compute and data described as fully self-controlled. The company says Spark now covers more than 130 languages, a claim positioned to support Chinese AI exports as a "second option" for the world.
That narrative travels well, but the model's standing in open benchmarks is more modest. Independent leaderboards place Spark outside the top tier on general capability, and iFlytek's chairman Liu Qingfeng has attributed gaps to limits of domestic accelerators on very long contexts — an honest framing that also doubles as a sovereignty argument.
Why the "department store" label sticks
Critics inside China call iFlytek an "AI general store" (AI杂货铺) — a firm that sells almost everything AI-adjacent yet leads in few categories. The pattern is real: shared speech and model foundations are repackaged into many hardware shells, but each front end needs its own sales, delivery and channel. When billions of RMB of R&D are divided across more than a dozen lines, no single product gets the concentrated firepower that pure-play rivals can muster.
Profit quality underscores the point. Government subsidies made up roughly 84%, 61%, 75% and 75% of net profit across 2022–2025 respectively; the 2025 annual report lists a 636 million RMB (about $88 million / HK$690 million) subsidy item. Stripping out such support, core operating profit remains slim after twenty-seven years of building.
Honest limitations
This article relies on iFlytek's own disclosures and Chinese media interpretations of its filings; segment "shares" and "first-place" claims are vendor or third-party market claims, not independently audited global rankings. Subsidy dependence is a structural fact from the annual report, but the precise counterfactual — how profitable the core business would be without it — is not disclosed. Benchmark positions shift quarterly and should be rechecked before any procurement decision. We have not modeled iFlytek's non-Chinese revenue, which remains a small fraction of the total.
What readers can do now
- Treat iFlytek as a one-stop supplier of speech, education and healthcare AI products rather than a frontier model vendor.
- If evaluating Spark for sovereign-AI deployments, request current benchmark evidence and confirm domestic-compute claims with a technical pilot.
- Track the 2026 interim report for whether subsidy dependence and per-line R&D intensity improve before committing to a multi-year contract.