iFlytek Spark V4.0

What it is

iFlytek Spark V4.0 (讯飞星火 V4.0) is a large language model released by the Chinese AI company iFlytek on 27 June 2024.^[1] It is the version the company put forward as reaching the level of OpenAI's GPT-4. Developers reach it through iFlytek's API, where it is served as "Spark4.0 Ultra" under the model name 4.0Ultra.^[3] A model of this kind takes a text prompt and produces a text response; the version number marks a generation within a family of models rather than a different kind of system.

A promise with a date on it

The launch was organised around a commitment the company had already made in public. Eight months before the release, iFlytek had said it would bring Spark to GPT-4 level in the first half of 2024; at the launch it said that promise had been met.^[2] Framing a release this way turns the model into a checkpoint on a published schedule as much as a product, and it invites comparison with the models it names.

Trained on a domestic platform

iFlytek said the model was trained entirely on a domestic computing platform co-built with Huawei, and presented that as self-sufficiency in the face of U.S. chip restrictions.^[2] Training at this scale means running very large numbers of accelerator cards together for weeks at a time, so where the hardware comes from decides whether a lab can keep increasing scale at all. iFlytek said the training run used the FlyingStar One cluster, which it describes as having 10,000 cards.^[1]

Benchmarks and the comparison with GPT-4 Turbo

iFlytek said the model surpassed 8 of 12 mainstream Chinese and English benchmarks, and ranked first across eight international mainstream test sets.^[1][2] Benchmarks of this kind are fixed sets of questions with known answers, used to compare models on tasks such as reading comprehension, reasoning and subject knowledge. They are useful because everyone can run the same set, but a score describes behaviour on that set of questions, not behaviour in open use.

Set against GPT-4 Turbo, iFlytek claimed better performance in five aspects while acknowledging that the model lagged in coding and multimodal abilities.^[2] Coding here means writing and working with program code; multimodal means handling inputs or outputs that go beyond text, such as images.

Xinhua reported that Spark V4.0 showed gains in complex instructions, complex logical reasoning, spatial reasoning and high-school mathematics.^[5] In one demonstration puzzle, the model inferred that the ball was on the bedroom floor.^[5] Spatial reasoning of that kind — working out where an object lies from a described scene — is a different task from retrieving a fact or continuing a sentence.

Personal space

Alongside the model, iFlytek added a "personal space" feature to the Spark app and desktop version. It tailors output using information users provide about their work, study and life.^[2] In practice this means the assistant can draw on standing context about a user instead of relying only on the text of the current question, which changes what has to be retyped in every conversation.

Where it sits in the line

At the launch, iFlytek chairman Liu Qingfeng said the gap with OpenAI would likely widen again, possibly beyond a year, once GPT-5 was released.^[2] That remark is a reminder that a model compared with a moving target is only level at one moment. The model itself was succeeded by Spark 4.0 Turbo, which iFlytek released in October 2024.^[4]