iFlytek releases Spark V4.0

The release

On 27 June 2024, iFlytek publicly released Spark V4.0, the 4.0 version of its Spark large model.^[1] The launch took place in Beijing and was presented by the company's chairman, Liu Qingfeng.^[3]

A model release of this kind is usually a package rather than a single file: a new version of the underlying model, plus changes to the products built on top of it. iFlytek's presentation covered both sides — benchmark results for the model itself, and new features in its consumer app.^[1][3]

Trained on a domestic cluster

The hardware was the detail iFlytek put forward most prominently. Spark V4.0 was trained on FlyingStar One, which the company describes as a domestic cluster of 10,000 cards.^[1]

Training a large language model means running the same enormous computation over a very large body of text, again and again, until the model's parameters settle into useful values. No single accelerator card is fast enough to do that work alone, so the job is spread across thousands of cards wired into one cluster, with the model's parameters divided among them. A cluster of 10,000 cards is a machine on that scale. Calling the cluster "domestic" means the accelerators were domestically produced rather than imported, so the training run did not rest on foreign hardware.^[1]

The benchmark claims

iFlytek said the model surpassed 8 of the 12 mainstream benchmarks it was measured on.^[1] The company also claimed that Spark V4.0 beat GPT-4 Turbo in five respects, while saying it still trailed in coding and multimodal ability.^[2]

A benchmark is a standardised set of tasks with a scoring rule, which lets models be compared on the same questions. Counting how many benchmarks a model "passes" is a coarse summary: it reports on how many test sets the model comes out ahead, not by what margin, and results also depend on how the prompts are written and how the evaluation is set up. The comparison with GPT-4 Turbo works as a reference point rather than a claim of overall parity — by iFlytek's own account, two important areas, writing code and handling images or speech together with text, remained behind.^[2]

The warning about the gap

Liu Qingfeng used the same presentation to warn that the gap with OpenAI could widen again after GPT-5.^[2] The remark frames model development as a moving frontier: naming GPT-4 Turbo as the yardstick says where the current version stands, and naming the next OpenAI model as a risk says that the distance is not fixed, because the reference itself keeps moving.

Alongside the model

Xinhua reported that a smart grading machine launched together with V4.0 was described by iFlytek as an industry first.^[3] Grading machines belong to the part of the business that applies models to school and exam work rather than to chat: the same recognition and scoring technology, packaged as a device. The Spark app and the desktop client, meanwhile, gained a feature the company calls "personal space".^[3]

What the release shows

Three things sit side by side in this launch. One is a training stack built on domestically produced accelerators, presented as the foundation of the model.^[1] The second is a set of self-reported comparisons against GPT-4 Turbo, stated with the shortfalls left in — coding and multimodal ability were named as areas still behind.^[2] The third is packaging: the model shipped together with hardware for grading and with a new feature in the app, so that the same release reached both developers and ordinary users.^[1][3]

The chairman's warning about GPT-5 is part of the same picture. A release that measures itself against a specific competitor's current model is a snapshot of a race, not a finishing line, and the people presenting it said as much.