The claim is the product

Most model launches lead with a benchmark. This one led with a supply chain. At the festival launch, Liu Qingfeng said Spark is the only model trained on domestic computing power and that every other model was trained on Nvidia GPUs. In IT Home's Chinese text: 「只有星火一家是用全国产算力训练的，其他全是在 N 卡上训练」. One Chinese report of the event added that other models, at most, were migrated to domestic compute arrays afterwards. In the context of US export controls on advanced chips, that sentence is the announcement: a Chinese lab saying, in public, that it no longer needs the hardware it is not allowed to buy freely.

The sentence is also difficult to reconcile with documentation that has been public for years, which is the subject of the next section. First, note what the launch itself did not supply. No launch report read for this article names the domestic chips Spark X2.6 was trained on. "Only" is a claim about every other lab's training clusters, including labs that do not disclose theirs. And no benchmark table for the 293B model was published in the launch reporting. The absence of a benchmark table does not bear on the training-hardware claim — a benchmark measures capability, not provenance — but it leaves the launch's code and agent claims without published evidence as well.

The Huawei counterexample

Huawei's own research repository documents PanGu-α, a family of large-scale autoregressive pretrained Chinese language models reaching 200 billion parameters. The repository states the models were developed under the MindSpore framework and trained on a cluster of Ascend 910 AI processors. The family's technical report, by researchers at Huawei Noah's Ark Lab, puts the cluster at 2,048 Ascend 910 processors. Ascend is Huawei's domestic AI processor line. PanGu-α alone is difficult to reconcile with a literal reading of Liu's statement: a large model, documented by its maker since 2021, that was not trained on Nvidia GPUs.

Nor does Huawei present PanGu-α as a lone historical case. At Huawei Connect in September 2026, Huawei stated that more than 40 models had been natively pre-trained on Ascend and CANN, its software stack — in the company's words, making Ascend the only proven domestic stack capable of model pre-training. Those are Huawei's figures, offered in its own ecosystem pitch, and they have not been independently audited. But a claim that every model outside Spark was trained on Nvidia GPUs cannot coexist, unqualified, with the other major domestic vendor's public record of dozens of models pre-trained on its own processors.

Liu may have intended a narrower comparison. The launch reporting does not define such a class, and this article will not invent one to rescue the claim. What can be stated precisely is narrower: iFlytek reports that Spark X2.6 itself was trained using domestic computing resources. The Huawei record does not challenge that provenance claim. It challenges the exclusivity wrapped around it.

What is actually specified

Underneath the claim, the specification is concrete, and it is iFlytek-reported throughout. Spark X2.6 is a mixture-of-experts model with 293 billion parameters in total and 30 billion active per token, and a context window extended to one million tokens. It is the first Spark model built on an end-cloud integrated architecture. iFlytek describes its code and agent capabilities as industry-leading. That phrase is the company's wording; the launch reporting published no benchmark results to set against it — no coding score, no agent evaluation, no comparison table. Readers are given the architecture and the adjective, and asked to connect them themselves.

The launch did describe engineering aimed at the known constraint of domestic chips, their memory bandwidth: a dynamic parallel training approach for long-context tasks that allocates compute by task length, and distributed inference over heterogeneous domestic clusters with a four-tier KV-cache design. Those techniques are described in the launch reporting without a reproducible baseline or independent evaluation, so their effect cannot be quantified from the published record.

The pricing, at least, is a number. iFlytek's limited-offer prices, announced on 22 October 2026: ¥0.9 per million input tokens, ¥0.08 per million on a cache hit, and ¥3.6 per million output tokens. Limited-offer prices are marketing instruments as much as price lists, and these should be read with the date attached. What they signal is intent: iFlytek wants developers to try the model while the domestic-compute story is in the news.

A separate, testable release

The launch also shipped a smaller model, and it needs to be kept separate from the claims above. Spark X2.6-4B is an on-device model released the same day. Its Hugging Face repository, published under the Spark team's XHToken account, contains model weights and configuration files in Hugging Face Transformers format, and the model card states that the Spark-X2.6 model series is licensed under the Apache 2.0 License. The card says the 4B retains exactly the same architecture as Spark-X2.5-4B, achieving its gains through improved post-training, and reports improvements in coding, agent, long-context, multilingual and reasoning measures — iFlytek's own comparison against its predecessor. It provides three configurable thinking levels: none, low and high. The launch reporting adds that iFlytek's own inference framework runs the model more efficiently than Ollama; the model card does not repeat that comparison, and it specifies no hardware, configuration or workload for it. No controlled test of the claim has been published.

The predecessor has a track record outsiders could at least observe in download counts: IT Home reported at the event that the X2.5-4B, open-sourced on 1 September 2026, ranks among the top four globally for edge-model downloads, without specifying the metric, platform or period.

What testing the 4B can establish is what the 4B can do. It cannot establish how the 293B model was trained, and it says nothing about any other lab's hardware. The four questions this launch raises have four different kinds of evidence. The 4B's capabilities can be tested against its published weights. The 293B model's code and agent performance would need reproducible benchmarks, and none were published. Its training provenance would need technical disclosure, hardware documentation or an independent audit, and none was provided. And the word "only" is answered by the documented record of other domestic-compute training efforts — where Huawei's PanGu-α and its Ascend pre-training statement already stand.

Sources

  1. [1] IT Home, report of the iFlytek Global 1024 Developer Festival launch (22 October 2026), read via KuCoin flash republication — the launch itself: Liu Qingfeng’s statement that Spark is the only model trained on domestic computing power while all others were trained on NVIDIA GPUs, the 293B-A30B MoE architecture and 1M context, the limited-offer pricing (¥0.9 input / ¥0.08 cache hit / ¥3.6 output per million tokens), the X2.6-4B open-source release with three new inference modes and the framework-versus-Ollama comparison, and the X2.5-4B top-four downloads report. Published 2026-10-22T08:04ZRead source
  2. [2] IT Home, launch report as republished by Sina Finance (22 October 2026, Chinese text) — the original wording of Liu’s statement (只有星火一家是用全国产算力训练的，其他全是在 N 卡上训练), the dynamic parallel training and distributed inference descriptions, the X2.6-4B improvements as iFlytek’s own comparison, the X2.5-4B top-four downloads report, and the limited-offer pricingRead source
  3. [3] Leikeji (Ray Technology) event report, republished by Sina Finance (22 October 2026, Chinese text) — a second account of Liu’s wording, adding that other models were at most migrated to domestic compute arrays after training; no comparison group is defined in this account eitherRead source
  4. [4] HiTechExpert, "iFlytek Unveils Spark X2.6" (22 October 2026, sourcing IT Home) — independent republication cross-check of the same launch facts: end-cloud architecture, domestic-platform training, the inference innovations, the X2.6-4B release, specifications and pricingRead source
  5. [5] Huawei Noah’s Ark Lab, Pretrained-Language-Model repository (GitHub, Huawei’s research repository, read first-hand) — documents PanGu-α as a large-scale autoregressive pretrained Chinese language model with up to 200B parameters, developed under MindSpore and trained on a cluster of Ascend 910 AI processorsRead source
  6. [6] Zeng et al. (Huawei Noah’s Ark Lab), "PanGu-α: Large-scale Autoregressive Pretrained Chinese Language Models with Auto-parallel Computation", arXiv:2104.12369 (2021), read first-hand — PanGu-α, up to 200 billion parameters, developed under MindSpore and trained on a cluster of 2,048 Ascend 910 AI processorsRead source
  7. [7] Huawei Connect 2026 keynote release text (David Wang keynote, September 2026), syndicated in full by StorageNewsletter and read first-hand — Huawei’s statement that over 40 models have been natively pre-trained on Ascend and CANN, "making it the only proven domestic stack capable of model pre-training"Read source
  8. [8] TechnoTime, "Huawei Accelerates AI Infrastructure Push With New Ascend Chips and Million-NPU SuperCluster" (22 September 2026) — reputable report of the same Huawei Connect 2026 statements, attributing the more-than-40-models figure to HuaweiRead source
  9. [9] SparkLLM Team (XHToken), "Spark-X2.6-4B" model card, Hugging Face (read first-hand) — the repository contains model weights and configuration files in Transformers format; the card states the same architecture as Spark-X2.5-4B with gains from improved post-training, three configurable thinking levels, and that the Spark-X2.6 model series is licensed under the Apache 2.0 License. The card does not repeat the launch reporting’s framework-versus-Ollama efficiency comparisonRead source
  10. [10] Longbridge, "iFLYTEK Q3 Revenue Grows Nearly 8%..." (20 October 2026) — pre-announcement confirmation, two days ahead, that iFlytek would release the new cloud-plus-edge Spark model on 22 October on a domestically produced computing-power foundationRead source