CONFIRMED: iFlytek launches Spark X2.5, trained end to end on Chinese chips

iFlytek officially launched Spark X2.5 this Monday, September 7, its new large language model built on a mixture of experts architecture with 293 billion total parameters and 30 billion active per inference call. The announcement came straight from the company, with simultaneous coverage from Chinese financial outlets including Sina Finance, Sohu and NetEase, and the model is already live on iFlytek's own developer platform, the Xinghuo Open Platform. The detail iFlytek chose to highlight is not the size, it is the origin: the company says the entire training and inference pipeline ran on domestic Chinese computing infrastructure, with no dependence on foreign chips.

Why training everything on domestic silicon changes the read

That detail matters more than it first appears. With US export controls limiting China's access to Nvidia's most advanced accelerators, any Chinese lab that can publicly claim a fully domestic training run becomes part of a larger story about compute sovereignty. MaxAssistant previously flagged a similar signal with the rumor that DeepSeek was negotiating for up to 160,000 Huawei Ascend chips to work around the same ceiling. Spark X2.5 fits that pattern: every announcement like this chips away, in practice and in messaging, at the argument that China would be paralyzed without Western hardware.

The numbers the company is willing to show, for now

iFlytek says the X2.5's focus was raising coding and agent capability, while also strengthening math and general comprehension. As of now, there is no independent benchmark comparing the model against rivals like Qwen3, DeepSeek V4 or GLM 5, nor a disclosed position on tables such as LMArena. The only numbers available so far come from the company itself, which is normal on launch day, but it means the model's actual competitiveness against China's more established models still depends on outside evaluation.

The path paved by the edge models a week earlier

The flagship launch did not arrive in isolation. On September 1, iFlytek had already open sourced two smaller versions, Spark X2.5-4B and X2.5-1.7B, both with a native context window of up to one million tokens, aimed at embedded use in vehicles, smart hardware and IoT devices. The sequence suggests a deliberate strategy: first claim ground at the edge with light, open models, then announce the closed flagship to compete head on in coding and agents.

One more name fighting for room behind Qwen, DeepSeek and Kimi

iFlytek is better known outside China for speech recognition than for frontier language models, which makes this launch something of a repositioning test. The company had already signaled, during its first half 2026 earnings call, its intent to launch a flagship general purpose model built entirely on domestic compute at this year's Global 1024 Developer Festival. X2.5 lands in the middle of an especially crowded release cycle, with OpenAI's GPT-6 Astra freshly out, xAI's Grok 4.7 promised for September 12, and rumors of a Claude Opus 5.1 circulating the same week.

What changes in MaxAssistant's read

Confirmation of a fully domestic training run is a concrete, verifiable claim backed by the company's own account, and it underscores how much the Chinese AI ecosystem is investing in proving independence from foreign hardware. But that is one thing, and actual model competitiveness is another. Without third party benchmarks, treating Spark X2.5 as a peer of Qwen3 or DeepSeek V4 would mean calling a verdict that does not exist yet. Anyone tracking the Chinese LLM market should note the name and wait for the first independent evaluation before drawing any weighty comparison.

Sources

PANews, iFlytek to Release Spark X2.5 General Purpose Large Model on September 7: https://panews.io/articles/01a055d1-0b16-74c5-8391-bceafa10112b | Sina Finance, iFlytek Spark X2.5 officially released: https://finance.sina.com.cn/stock/t/2026-09-07/doc-iniqyriy6377862.shtml | TMTPost, iFlytek to Release Spark X2.5 Base Model on Sept. 7: https://en.tmtpost.com/news/8122532