As large-model applications spread, data security, high token costs, and insufficient computing power for complex tasks have become real-world bottlenecks to deployment.

Iflytek's answer is an edge-cloud collaborative allocation scheme that sorts tasks into routine and complex categories, handing the former to on-device models and the latter to the cloud, in an attempt to balance data security, cost control, and capability ceiling.

To underpin that architecture, Iflytek also released its Spark X2.6 large model along with an open-source on-device model, and rolled out AI applications for education, healthcare, and enterprise office work.

Three Bottlenecks: Data That Can't Go to the Cloud, Tokens That Cost Too Much, and Complex Tasks the Edge Can't Handle

The 2026 Iflytek Global 1024 Developer Festival opened in Beijing on October 22. Speaking at the main forum, Iflytek Chairman Liu Qingfeng said data security, model capability, token costs, and organizational effectiveness have all become points of concern for enterprises.

"AI today is making 80% of individuals more productive, yet only 6% of enterprises see a clear productivity gain," Liu said.

He grouped the obstacles to enterprise-scale large-model deployment into three categories.

First, companies are unwilling to put sensitive data in the cloud. Core business data touches privacy and compliance red lines, and enterprises have serious security concerns about letting it leave their own boundaries.

Second, complex tasks still need the cloud. Work that demands high reasoning ability and strong logic depends on cloud computing power; on-device processing alone cannot deliver it.

Third, token consumption climbs sharply as applications go deeper, turning steep API fees into a heavy burden on rolling out a business.

Finding the optimal balance between security, cost, and capability has become the key to moving large models from technology into industry.

60% Local, 30% Collaborative, 10% Cloud

To address those pain points, Liu laid out Iflytek's edge-cloud collaborative allocation scheme.

The scheme divides work by task complexity:

60% of routine applications run on-device, keeping data within the organization at zero token cost for inference;

30% of tasks use an edge-cloud collaborative mode, which the company says cuts token costs by more than fivefold while still protecting privacy;

The remaining 10%, highly complex and compute-intensive tasks, go to the cloud-based large model.

By Iflytek's account, the approach lowers token consumption and spares enterprises from pushing all their data to the cloud, while preserving the capability ceiling needed for hard problems.

Spark X2.6: 293B-A30B Parameters, 1M Context

Iflytek released the Spark X2.6 large model the same day as the core support for its edge-cloud architecture.

"Spark X2.6 is the only model among the mainstream large models on the market today trained on domestic computing power," Liu said. As a fully self-controlled, full-stack model, it required Iflytek to be the first to break through the training-efficiency bottleneck domestic chips face on long-horizon tasks.

The model uses a 293B-A30B MoE architecture, with context length extended to 1M.

On pricing, Spark X2.6's cloud API carries promotional rates of 0.9 yuan per million input tokens, 0.08 yuan per million tokens for cached input hits, and 3.6 yuan per million output tokens. Inference on the open-source Spark X2.6-4B on-device model, released at the same time, is free for a limited period to lower the barrier to edge deployment.

Education, Healthcare, and Office Work Come First

Built on the integrated edge-cloud architecture, Iflytek also unveiled new AI products for several industries.

In education, it launched the Iflytek Starlight super-agent workbench, which draws on an education-specific knowledge base to assist teachers, and added an "inquiry-based learning" mode to its AI learning devices.

In healthcare, the new Iflytek Xiaoyi extends from one-off consultations to continuous health management, adding integrated interpretation of multiple medical reports, health risk insights, and traditional Chinese medicine tongue and facial constitution assessment.

For enterprise office work, it released the Xingfan WinWork digital employee platform, an effort to weave digital employees into corporate business processes.

For the past three years, the AI industry has competed largely on model parameters, leaderboard scores, and cloud computing power. In 2026, the center of gravity for how enterprises evaluate AI has shifted to running it continuously at a cost they can bear, data boundaries, and organizational effectiveness.

The "60% handled locally" claim in Iflytek's edge-cloud scheme still needs to be tested across different industries, devices, and levels of business complexity. But it does offer a relatively pragmatic path for large models to move from demo-grade capability to production-grade applications.

At the very least, it means enterprises facing large models no longer have to pick just one of security, cost, and intelligence.