iFlytek's Spark X2.6 Puts Fully Domestic Compute at the Center of China's Sovereign AI Push

Spark X2.6 is a 293-billion-parameter mixture-of-experts large language model from iFlytek that was launched Oct. 22, 2026, with a 1 million token context window and training carried out entirely on domestic Chinese computing infrastructure.

iFlytek released Spark X2.6 on Oct. 22, 2026, at the Global 1024 Developer Festival in Beijing and simultaneously open-sourced the Spark X2.6-4B on-device model. ifeng.com and IT Home reported that the company called the release the first time it achieved breakthroughs in coding and agentic capabilities using fully domestic computing power. People's Daily Online described Spark X2.6 as an end-cloud integrated LLM, meaning the same model family is intended to run both in data centers and on end-user devices.

The launch is part of a broader pattern in which Chinese AI vendors are pairing large models with domestic infrastructure claims. People's Daily Online's report, carried by Global Times, said Spark X2.6 was trained entirely on domestic Chinese computing infrastructure and deployed on a domestic 10,000-card inference cluster. That sentence is the core of the sovereign AI narrative around the product.

What is new in Spark X2.6's architecture?

Spark X2.6 uses a mixture-of-experts architecture with 293 billion total parameters and 30 billion activated per token. In an MoE model, the total parameter count represents the full set of expert weights, while the activated count is the number of parameters used to process each token. The 30B activation figure determines compute cost at inference time and is the reason a 1 million token context window is operationally plausible.

The 1 million token context window lets Spark X2.6 ingest the equivalent of multiple long technical documents or a large code repository in a single pass. For coding and agentic workloads, that changes how much state the model can hold before it needs external tools or summarization. iFlytek's decision to emphasize coding and agentic capabilities, as reported by ifeng.com, suggests the company is targeting developer tooling and autonomous task execution rather than general chat.

People's Daily Online attributed the scale figures to the launch event, and the numbers are consistent with the current frontier of large MoE models. The more contested claim is the training infrastructure: the report says the model was trained entirely on domestic Chinese computing infrastructure, a statement that is difficult to verify from outside and that iFlytek did not support with published training details.

How did iFlytek train the model on fully domestic computing power?

Liu Qingfeng, chairperson of iFlytek, used the launch to make a policy-relevant efficiency claim. “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,” he said, according to People's Daily Online. The word “lead” asserts that iFlytek is ahead of domestic rivals on this specific problem, not just that the model works.

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.Liu Qingfeng, Chairperson of iFlytek

The efficiency bottleneck Liu referred to is the gap between the compute utilization that frontier models achieve on leading foreign accelerators and what is possible on domestic alternatives. Long-horizon tasks, such as training runs that span many steps or agentic loops that chain many model calls, are where that gap shows up most clearly. iFlytek's claim is that Spark X2.6 closes enough of that gap to make domestic chips viable for production-scale training and inference.

What does the open-source Spark X2.6-4B add?

The Spark X2.6-4B release is a separate product decision with a different audience. BigGo Finance reported that iFlytek described the open-source model as a step change in on-device agent and coding capabilities. On-device models are attractive to enterprises that want low latency, offline operation, or strict data residency, because the weights run on the customer's hardware rather than on iFlytek's cloud.

The 4B parameter size is small enough to run on current-generation laptops, phones and edge servers, which is the segment iFlytek is targeting with the on-device release. The open-source decision also lowers the barrier to entry for developers who want to fine-tune the model for specific coding or agent tasks without paying for cloud inference.

The device-cloud collaborative architecture announced at the same event is the glue between the two models. The architecture is designed to route tasks between device, cloud, or hybrid execution, per BigGo Finance's report. iFlytek did not disclose the routing policy details in the launch coverage, but the architectural pattern is recognizable from other vendors' hybrid inference offerings.

What technical details matter for deployment?

Three technical decisions in the Spark X2.6 launch matter more than the headline numbers. The first is the activation ratio: at 30 billion activated parameters out of 293 billion total, roughly 10% of the model is active per token, which is the main lever for keeping inference cost below what a dense model of the same size would require.

The second is the context window implementation. A 1 million token window requires an efficient attention mechanism, because full attention over 1 million tokens is computationally prohibitive. iFlytek did not disclose which attention variant Spark X2.6 uses, so deployment teams should test long-context recall rather than assuming the window is fully usable.

The third is the device-cloud routing boundary. The architecture splits tasks between device and cloud, but the launch coverage does not specify the criteria the router uses or whether customers can enforce a strict no-cloud policy for designated workloads.

1. Whether the 1 million token context window is available at the same price tier as the standard API, because long-context pricing can change the economics of agent workloads.
2. Whether the open-source Spark X2.6-4B license permits commercial redistribution, since on-device deployments often require embedding the model in a product.
3. Whether the device-cloud routing architecture can be configured to keep specified workloads off the cloud entirely, for data-sovereignty reasons.
4. Whether the limited-time API discount has a stated end date, because the 40% reduction changes the cost per token.

How is iFlytek pricing Spark X2.6?

BigGo Finance published the limited-time API pricing for Spark X2.6: 0.9 yuan per million input tokens, 3.6 yuan per million output tokens, and 0.08 yuan per million tokens on cache hits. The report said these prices reflect a 40% discount. At those rates, a 1 million token input costs 0.9 yuan, and a 1 million token output costs 3.6 yuan.

Pricing tier	Price per million tokens	Notes
Input	0.9 yuan	Limited-time, 40% discount
Output	3.6 yuan	Limited-time, 40% discount
Cache hit	0.08 yuan	Limited-time, 40% discount

The cache-hit price is the most distinctive number in the schedule. Cache hits occur when a model reuses previously processed context, such as a system prompt or a long document that is sent with every request. Pricing cache hits at 0.08 yuan per million tokens gives developers an incentive to structure agentic workloads so that stable context is cached rather than reprocessed.

What does this mean for China's sovereign AI push?

The sovereign AI framing is explicit in the launch. iFlytek is marketing Spark X2.6 as a model that can be trained and served without foreign accelerators, which aligns with the policy goal Beijing has pushed since U.S. export controls on advanced chips were tightened. People's Daily Online's coverage validates that framing by describing the model as “based on fully domestic computing power.”

The domestic 10,000-card inference cluster cited by People's Daily Online is a scale figure that also signals supply. A cluster of that size is large enough to serve a popular model at production volume, and iFlytek's willingness to state the number suggests it expects the claim to be checked.

For enterprise buyers inside China, the domestic compute claim can be a procurement advantage, because state-owned and regulated customers often weigh technology self-reliance in purchasing decisions. For buyers outside China, the same claim is a governance signal: the model's development is tied to a national technology strategy, and the vendor has not published independent evidence for the training infrastructure claim.

How are experts and the market reacting?

The most concrete reaction in the source material is Liu Qingfeng's own statement, which People's Daily Online quoted. No independent benchmarks, third-party evaluations, or rival vendor comments were included in the launch coverage reviewed for this article. That leaves the efficiency and capability claims resting on iFlytek's own presentation.

What should enterprises watch next?

Enterprise teams that want to evaluate Spark X2.6 should start by measuring the 1 million token context window on their own data, because context length is only useful when the model can retrieve and reason over the content inside it. The second test is the cache-hit rate in a real workload, since the 0.08 yuan price only matters if the routing architecture actually reuses context across calls.

The open-source Spark X2.6-4B gives engineering teams a low-cost way to test the family without signing up for the cloud API. If the 4B model performs well on device for coding and agent tasks, it could become the entry point that pulls developers into the broader Spark ecosystem and the discounted cloud pricing.

The pricing discount is labeled limited-time, which means enterprise buyers who standardize on Spark X2.6 now should model what the API will cost after the promotion ends. The 40% figure from BigGo Finance is the only pricing data in the launch coverage, and there is no published timeline for when the discount expires.

iFlytek's launch is a reference point for where the Chinese LLM market is in late 2026: large MoE architectures, million-token contexts, open-source edge models, and a domestic compute claim are now bundled into a single product announcement. The question for enterprises is whether the claims survive independent testing, and whether the pricing and routing architecture hold up in production.