Most Chinese large models (大模型) are chasing the chatbot crown. iFlytek (科大讯飞), the Hefei company that has spent two decades on speech recognition, aimed its model at two quieter markets: the schoolroom and the phone call.
Key takeaways
- iFlytek released the first version of Spark (星火) on 6 May 2023, and by January 2024 had trained a major upgrade on fully domestic compute.
- Spark V3.5, launched 30 January 2024, was the first widely deployed Chinese large model trained on a home-grown "飞星一号" (Feixing-1) compute platform built with Huawei.
- On the same day iFlytek released a speech large model (语音大模型) that beat OpenAI's Whisper V3 on its first 37 mainstream languages.
- An open-source Spark-13B checkpoint, adapted to domestic hardware, shipped the same week.
- Spark's teaching assistant had reached 584 schools in Guangdong province by May 2024, showing a deployment path that is education-first rather than consumer-first.
A model shaped by voice
iFlytek is not a typical model lab. Founded in 1999, it built its reputation on speech—dictation, transcription, and later real-time translation devices. When it entered the large-model race, that history bent the product. Spark's differentiator is not raw reasoning leaderboard scores but the combination of a text large model with a speech large model that understands and speaks.
At the 30 January 2024 event, iFlytek unveiled the Spark speech large model (星火语音大模型). On its first 37 mainstream languages, the company reported recognition quality exceeding OpenAI's Whisper V3, with synthetic speech rated above 83% on a human-likeness measure across 40 languages. For a company whose flagship hardware is a translator, that is the feature that matters most.
Domestic compute as a constraint and a selling point
The same launch introduced Spark V3.5, described by iFlytek as the first broadly available Chinese large model trained entirely on domestic compute. The "飞星一号" (Feixing-1) platform, a ten-thousand-card domestic cluster co-built with Huawei, had been switched on in October 2023. Within roughly 90 days, iFlytek used it to train a model whose math, language understanding and voice interaction were said to exceed GPT-4 Turbo on several dimensions.
Training on domestic hardware is both a strategic hedge against export controls and a marketing claim aimed at government and enterprise buyers who must satisfy localisation requirements. It also forces the same efficiency discipline DeepSeek later became famous for: you optimise when you cannot simply buy more.
An education-first deployment
Where many labs lead with a consumer app, iFlytek led with classrooms. Its Spark teaching assistant (星火教师助手) was deployed across 584 schools in Guangdong by May 2024, supporting lesson preparation, essay review and spoken-English practice. The company also shipped a "smart blackboard" and, in October 2024, a high-school math tutor built on a "question chain" (问题链) method. iFlytek said its 4.0 Turbo release that month took first place in 9 of 14 international evaluations, with math problem-solving surpassing GPT-4o.
This is a defensible niche. Education is regulated, procurement-driven, and language-specific—exactly the setting where a domestic, speech-fluent model with local deployment options wins against general-purpose foreign models that may not be approved, localised, or affordable at scale.
Why voice is the moat
Speech is where iFlytek's two decades of data pay off. China's market spans dozens of regional dialects and constant code-switching between Mandarin and local languages, a setting where a generic text model paired with an off-the-shelf speech engine tends to underperform. iFlytek's edge is the integrated stack: its own acoustic models, the Spark text model, and a speech large model trained to attach language, content, prosody, timbre and emotion to a single utterance. For call-centre, in-car and robot interactions—scenarios the company named at the launch—that integration, not a leaderboard score, is what actually ships.
That logic also explains the open-weight move. The Spark-13B checkpoint, released the same week and adapted to the Ascend (昇腾) and MindSpore (昇思) software stack, lets domestic hardware partners and academic researchers build on a model already tuned for Chinese speech. It extends iFlytek's reach beyond its own cloud without asking adopters to rebuild the acoustic layer from scratch.
Honest limitations
The benchmark and deployment figures here come from iFlytek's own launch materials and provincial education reports; we have not independently re-run the model or audited the 584-school count beyond the company's stated coverage. "Surpassing GPT-4 Turbo" is a per-capability claim from early 2024 and does not reflect either model's current version. Spark's global footprint is far smaller than the consumer chatbots, and its advantage is concentrated in Chinese-language education and speech rather than general reasoning.
What readers can do now
- If you build for Chinese education or public-sector deployment, evaluate Spark alongside the general models—local compute compatibility and speech are genuine differentiators.
- For multilingual speech products, test the Spark speech model on your target languages rather than assuming Whisper leads by default.
- When a vendor cites "domestic compute," ask which stack (chip, framework, cluster) and whether inference is certified for your compliance regime.