iFlytek

Artificial Intelligence · Foundation Models · China Tech

Launch

iFlytek launches Spark X2.6, a 293B hybrid cloud-edge AI model

A claim that the model trained purely on domestic chips tests how far US export curbs have pushed Chinese AI self-sufficiency.

• iFlytek unveiled Spark X2.6, a 293-billion-parameter mixture-of-experts model with a hybrid cloud-edge architecture, at its 1024 Global Developer Festival in Beijing on October 22, 2026.
• iFlytek, a Hefei-based Chinese AI company best known for speech recognition, built Spark X2.6 as its newest flagship model targeting coding and autonomous-agent tasks.
• The model activates about 30 billion of its 293 billion parameters per token and supports a 1-million-token context window, splitting simple tasks on-device and complex ones in the cloud.
• iFlytek open-sourced a compact Spark-X2.6-4B model under Apache 2.0 and a 1.7B on-device agent variant with day-0 support in vLLM, including NPU deployment via vLLM-Ascend.
• iFlytek says Spark X2.6 was trained entirely on Chinese-made chips and prices API access at 0.9 yuan per million input tokens and 3.6 yuan per million output tokens.
• iFlytek gave no chip list, training-compute figures, or flagship benchmark results to support its domestic-chip claim, which remains independently unverified.
• The launch signals Chinese AI labs are racing to prove they can match coding and agent capabilities without relying on foreign chips, a key test of self-sufficiency amid US export controls.