iFlytek's Spark X2.6 is a 293B AI model trained entirely on Chinese chips — no NVIDIA in sight

• Who: iFlytek — announced October 22, 2026, at the iFlytek Global 1024 Developer Festival, by chairman Liu Qingfeng.
• What: Spark X2.6, a 293-billion-parameter mixture-of-experts model (30B active per query, A30B) with a 1-million-token context window and an end-to-cloud integrated architecture.
• The silicon claim: trained entirely on domestic Chinese computing resources — Liu said Spark is the only model trained on domestic compute, with all others trained on NVIDIA GPUs, and the smallest of its kind on domestic infrastructure.
• Engineering: dynamic parallel training to break long-context training bottlenecks on domestic chips, plus distributed inference over domestic heterogeneous compute — targeting memory-bandwidth limits of home-grown chips.
• Open source: Spark X2.6-4B edge model, beating its X2.5-4B predecessor on all metrics; the X2.5-4B line has been in the top 4 edge open-source models by downloads since September 1. Three new inference modes; free for a limited time.
• Price: input 0.9 yuan per million tokens, input cache hit 0.08 yuan, output 3.6 yuan (limited offer); API trial open on iFlytek's official site.

iFlytek picked its annual 1024 Developer Festival on October 22 to launch Spark X2.6, the newest generation of its large language model. The specs are heavyweight — 293 billion parameters, a million-token context window — but the real story is the silicon: the company says the model was trained entirely on domestically produced Chinese computing hardware, with no NVIDIA GPUs in the pipeline.

Chairman Liu Qingfeng framed it as an industry first on stage: Spark, he claimed, is the only large model trained on domestic computing power, while every competitor's model was trained on NVIDIA GPUs — and among domestic-compute-trained models, Spark is the smallest in size. Those are company claims, but they land in the middle of the most important hardware question in AI: can frontier-class models be built on China's home-grown chips?

The model

Spark X2.6 uses a mixture-of-experts architecture with 293 billion total parameters and 30 billion activated per query (A30B), and supports up to 1 million tokens of context. iFlytek describes the architecture as end-to-cloud integrated, aimed at the two workloads the company wants to own: code generation and agentic task execution, where it claims industry-leading capability.

To get long-context training working on domestic chips, iFlytek says it broke through a training-efficiency bottleneck with a dynamic parallel training technique tuned for domestic platforms — addressing the low mixed-training efficiency that comes from the limited VRAM bandwidth of home-grown chips. On the inference side, it introduced distributed inference over domestic heterogeneous computing resources and result reuse that skips recalculation, both aimed at the same bandwidth problem.

The small one goes open source

Launching alongside the flagship is Spark X2.6-4B, a lightweight open-source model aimed at edge devices. iFlytek says it improves on every metric over its predecessor, X2.5-4B — and that predecessor has a track record: since its release on September 1, it has ranked among the top four edge-device open-source models globally by downloads.

The 4B model gets three new inference modes, and iFlytek claims its inference efficiency on its own stack beats Ollama. It is free for a limited time — an obvious play to seed developer adoption before the meter starts running.

The price card

iFlytek published API pricing for the full X2.6 model, marked as a limited offer: 0.9 yuan per million input tokens, 0.08 yuan per million on input cache hits, and 3.6 yuan per million output tokens. The model is available for trial through iFlytek's official channels, with professional security options and customizable policies for specific task scenarios, per the launch coverage.

Why it matters

This is the strongest data point yet for the "can China do it without NVIDIA" thesis. If a 293B-parameter MoE with million-token context and top-tier coding and agent scores really did come off a domestic training stack, the gap between US-controlled accelerators and Chinese alternatives is narrowing faster than most roadmaps assumed — which is exactly what the US export-control regime was designed to prevent.

The caveats matter too. "Industry's highest level" for coding and agent capabilities is the company's own characterization of launch-day results, and the day-one coverage did not include independent benchmark tables. And the pricing — sub-yuan per million input tokens — signals that iFlytek intends to compete the same way Chinese labs usually do: on price. (IT Home via KuCoin; China AI briefing)