iFlytek Spark 4.0 Turbo

What it is

iFlytek Spark 4.0 Turbo is a version of the Spark large language model from the Chinese company iFlytek. It was released on 24 October 2024 at the 7th World Voice Expo and the 2024 iFlytek Global 1024 Developer Festival, where chairman Liu Qingfeng presented it.^[1] It is a general model — it works with text, images and documents rather than voice alone — and it is aimed at developers building applications as well as enterprises running it in-house.

Growing out of a voice model

Spark's earlier strengths were in voice. Far-field audio lets a device pick up speech from across a room; full-duplex interaction means the system can listen and speak at the same time instead of waiting for its turn; multilingual support covers more than one language. The 4.0 Turbo update kept all three and added multimodal input, "super anthropomorphic" interaction and personalisation.^[1] Multimodal means the model reads images and page layouts as well as plain text. "Super anthropomorphic" refers to making the spoken voice sound closer to a person, and personalisation to adapting responses to the individual user. The direction is a model you talk to rather than type at.

The comparison iFlytek published

iFlytek said that in back-to-back tests on fresh data the model's seven core capabilities comprehensively surpassed GPT-4 Turbo, and that its mathematics and coding were better than GPT-4o.^[1] It also said the model placed first in 9 of 14 mainstream Chinese and English benchmarks.^[1] Those are the company's own comparisons, run on test data it describes as fresh; the benchmarks and the scoring are set by iFlytek rather than by an outside party.

The computing base

The model was trained and run on Feixing No.1 (飞星一号), a domestic computing platform; Feixing No.2 was announced at the same event.^[1] iFlytek spent more than a year optimising training and inference on the 10,000-card Feixing No.1 system.^[1] "Domestic" here means the chips and the software stack come from inside China rather than from the imported accelerators most large models have been trained on, and moving a model onto unfamiliar hardware usually means rewriting low-level routines, tuning how work is split across cards, and re-checking numerical behaviour. Liu Cong acknowledged that iFlytek's computing scale is smaller than that of leading companies and that domestic platforms require considerable adaptation work.^[1]

The January 2025 upgrade

On 15 January 2025 the base model was upgraded again, this time benchmarked against the then-latest GPT-4o.^[2] Mathematics was improved with synthetic data generated by Spark X1; synthetic data here means training examples written by a model rather than by people.^[2] Accuracy on document layout and OCR (optical character recognition) rose by 40%, and on complex tables and handwriting by 60%.^[2] The same upgrade added an agent-based framework for long texts with sentence-level source tracing, which cut errors on knowledge questions by 40%.^[2] An agent-based framework splits the reading of a long document into steps that the model works through one at a time; sentence-level tracing ties each sentence of an answer back to the passage it came from, so a reader can check it.

Where it sits

iFlytek's Spark line is not only a consumer product. As of October 2024 the company said it had co-built more than 20 industry large models with leading enterprises, covering over 300 application scenarios.^[1] Spark 4.0 Turbo is the general-purpose base in that arrangement: one model supplies the shared capability, and industry models are trained on top of it for particular businesses. The announcement of Feixing No.2 at the same event points to the next stage of the computing build-out that the model depends on.^[1]