iFlytek Says Spark X2.6 Trained on Chinese Chips

Chinese AI company iFlytek has released Spark X2.6, a large language model it says was trained entirely on Chinese-made chips, at its 1024 Global Developer Festival in Beijing on Thursday. The model uses a mixture-of-experts design with 293bn total parameters, 30bn of them active for each token, supports a context window of 1mn tokens and runs on “a domestic 10,000-card inference cluster”, Global Times reported.

“Spark-X2.6 has taken the lead in breaking through the efficiency bottleneck of long-horizon task training on domestic chips, truly achieving independence and controllability with homegrown computing power,” iFlytek chairman Liu Qingfeng said at the event.

iFlytek also released a 4bn-parameter version for phones and other devices, and its weights are public on Hugging Face under the Apache 2.0 licence. The flagship’s weights are not. According to IT Home, via Sina, the X2.6 series is available to try on iFlytek’s Spark website.

Liu says only Spark was trained on fully domestic computing power

Liu went further on stage. He said Spark is the only model trained on fully domestic computing power, while the others were all trained on “N cards”, IT Home reported. He added that iFlytek’s model is also the smallest of them. TSN has not verified either claim, and iFlytek did not name the rival models.

According to IT Home, iFlytek says it developed a dynamic parallel training technique for long tasks, adapted to domestic hardware, to deal with the low efficiency of mixed training caused by the limited memory bandwidth of Chinese chips. Neither report says which chips were used or how many trained the model; the 10,000-card figure Global Times gives is for inference.

iFlytek published no benchmark scores for the flagship

iFlytek says Spark X2.6 performs at a level “comparable to leading industry systems in multiple technical benchmarks”, Global Times reported. It did not publish the figures in the reports TSN checked, and Liu argued that the model’s performance in real deployments, in state-owned enterprises, education, healthcare and the courts, is better than its public benchmark scores suggest.

The smaller model comes with more detail. Its Hugging Face model card says Spark-X2.6-4B “retains exactly the same architecture as Spark-X2.5-4B”, with its gains coming from improved post-training, and offers three thinking levels: none, low and high. The repository lists about 4.1bn parameters. The model card gives setup instructions for both Nvidia GPUs and Huawei’s Ascend NPUs.

The weights went up before the event. The SparkLLM account announced the 4B model on X on Tuesday evening and posted self-reported scores, including 71.8 on BFCL-V4 and 54.3 on SWE Pro, in a follow-up. Those are the company’s own results and have not been independently checked. IT Home says the previous X2.5-4B model, open-sourced on 1 September, has entered the global top four on-device models by downloads, also a company claim.

iFlytek has been on the US Entity List since 2019

The US Commerce Department added iFlytek to its Entity List in October 2019, alongside Hikvision, SenseTime and others it said were “implicated in human rights violations and abuses” in Xinjiang. The listing requires a licence for exports of all items subject to US export rules to the company. iFlytek is still on the current Entity List.

The listing does not prove or disprove the chip claim. iFlytek has not named the chips, the size of the training cluster or the compute used, and has not published a technical report. Liu’s statement that rivals train on “N cards” is his own. This week, Reuters reported that ByteDance, Alibaba and Tencent had ordered more than 3mn Nvidia RTX PRO 5500 chips, while DeepSeek has open-sourced its own inference engine. Global Times also cited domestic reports that a Shenzhen team used Huawei Ascend 910C chips to post-train DeepSeek-V4-Pro.

What developers can test now is the 4B model, which joins other recent open releases tracked by TSN. The 293B flagship’s training claim will stay unverifiable until iFlytek names the hardware.