Home › News › AI Daily October 25 Evening Edition

AI Daily | October 25, 2026 Evening Edition: iFlytek Launches SparkAudio 1.0, Nearly Half of South Korea's New AI Models Built on Chinese Foundation Models

Published: 2026-10-25 20:00  |  Category: AI Daily  |  Read: about 8 min

Top Stories This Evening

  • iFlytek launches the SparkAudio 1.0 voice foundation model: device-cloud architecture, 98.7% accuracy, and a 91.4% comprehension rate even at 85dB of noise
  • Nearly half of South Korea's new AI models are built on Chinese foundation models: according to The Korea Times, China's underlying AI technology is accelerating its penetration into overseas markets
  • The MLEIL2026 international conference opens in Changsha: focusing on frontier research in machine learning, embodied intelligence and large models
  • Shanghai issues an action plan for AI-enabled industrial services: pushing AI from the consumer side deep into the industrial side
  • The first Shanghai "AI + Consumption" Innovation Competition holds its awards: more than 800 projects entered, accelerating AI adoption in consumer scenarios

1. iFlytek Launches SparkAudio 1.0 Voice Foundation Model with 98.7% Accuracy

iFlytekSparkAudioVoice Foundation Model

On October 25, iFlytek officially launched the SparkAudio 1.0 voice foundation model at the Hefei Binhu International Convention and Exhibition Center; it is the "auditory nerve center" embedded in the iFlytek Spark X2.6 large-model system.

  • Adopts a device-cloud architecture with on-device response latency below 180 milliseconds
  • Speech recognition accuracy reaches 98.7%, setting an industry record
  • Maintains a 91.4% semantic comprehension rate even in 85dB high-noise environments
  • Supports a privacy mode: sensitive voice data can be processed entirely on-device and never sent to the cloud
  • Deeply integrated with Spark's embodied multimodal interaction technology to empower smart hardware scenarios
Editor's note: The biggest highlight of SparkAudio 1.0 is not accuracy itself but its "noise as context" design philosophy: a 91.4% comprehension rate at 885dB of noise means AI can finally "hear clearly" in real scenarios such as factory floors and noisy streets. The device-cloud architecture balances privacy and performance, clearing the auditory obstacle for the deployment of embodied intelligence.

2. Nearly Half of South Korea's New AI Models Are Built on Chinese Foundation Models

South KoreaChinese AIGlobalization

According to a report by The Korea Times on October 25, data shows that nearly half of the AI models newly released by local South Korean developers this year are built on Chinese foundation models, drawing widespread industry attention.

  • South Korean developers' reliance on Chinese foundation models has climbed sharply over the past year
  • The main Chinese models used include Qwen, DeepSeek and other series
  • The report notes that the phenomenon has triggered debate over "technological sovereignty"
  • The advantages of Chinese AI models in cost-effectiveness and multilingual capability are the main draws
Industry signal: Chinese large models are moving from "domestic leadership" toward "global infrastructure." The fact that South Korea, a technology powerhouse, still adopts Chinese foundation models in large numbers shows that domestic models' competitiveness in the open-source ecosystem has reached a new height.

3. The MLEIL2026 International Conference Opens in Changsha

MLEIL2026Academic ConferenceEmbodied Intelligence

From October 23 to 25, the 2026 International Conference on Machine Learning, Embodied Intelligence and Large Models (MLEIL2026) was held in Changsha, co-organized by Hunan University of Information Technology and other institutions.

  • Focuses on three major directions: foundational theory of large models, embodied intelligence and multimodal learning
  • Experts from Huawei, iFlytek, the Shanghai AI Laboratory and other companies and institutions shared technical insights
  • Discussed topics such as AI-native corpus development, industry large-model training and large-model evaluation systems
  • Showcased the latest progress of artificial intelligence from basic research to industrial deployment

4. Shanghai Issues an Action Plan for AI-Empowered Industrial Services

ShanghaiIndustrial PolicyAI + Industry

Shanghai has officially issued an action plan for AI-enabled industrial services, pushing artificial intelligence from the consumer side deep into the industrial side.

  • Aims for a significant increase in AI penetration across manufacturing, logistics and finance by 2027
  • Focuses on supporting industrial large models, intelligent quality inspection and predictive maintenance
  • Encourages companies to build AI-native industrial software and data platforms
  • Includes supporting funds and the construction of technology incubation platforms
Policy direction: Following "AI + consumption," "AI + industry" is becoming the next policy focus. Shanghai's plan sends a clear signal: AI cannot stop at the chatbot stage; it must penetrate the core links of manufacturing.

5. The First Shanghai "AI + Consumption" Innovation Competition Holds Its Awards

ShanghaiAI + ConsumptionInnovation Competition

On the morning of October 25, the award ceremony of the first Shanghai "AI + Consumption" Innovation Competition was held at the West Bund International Technology and Consumer Carnival (WTCC).

  • The competition attracted more than 800 high-quality projects
  • Covering multiple consumption scenarios such as smart retail, smart dining and personalized recommendations
  • Taobao's Double 11 search upgraded with multimodal intent recognition, using the Qwen large model to uniformly understand images, text, audio and video
  • AI is moving from "technology showcase" to landing in real consumption scenarios

6. Taobao's Double 11 Search Upgrade: Multimodal Intent Recognition Goes Fully Live

TaobaoDouble 11Multimodal Search

With Double 11 approaching, Taobao has announced a full upgrade of its search system, achieving multimodal intent recognition based on the Qwen large model.

  • Moving beyond traditional keyword matching to uniformly understand images, text, audio and video inputs
  • Can automatically break down vague needs such as "gifting" and "outfit matching" to deliver precise recommendations
  • Users can upload images or describe things by voice for natural-language-style search
  • One of the largest-scale deployments of large models in core e-commerce scenarios

7. The Ministry of Commerce Revises the AI Export Control List

Ministry of CommerceExport ControlsAI Governance

The Ministry of Commerce and the Ministry of Science and Technology recently issued Announcement No. 48 of 2026, seeking public comment on the revision of the Catalogue of Technologies Prohibited or Restricted from Export in China.

  • Adjustments to export control provisions concerning AI-related technologies
  • Reflects the latest changes in the global AI technology competition landscape
  • May affect domestic AI companies' overseas business arrangements

Today's Data at a Glance

MetricFigure
SparkAudio 1.0 speech recognition accuracy98.7%
SparkAudio 1.0 comprehension rate in noisy environments (85dB)91.4%
SparkAudio 1.0 on-device response latency<180ms
Share of South Korea's new AI models built on Chinese foundation modelsNearly 50%
Entries in the Shanghai AI + Consumption competition800+
Data types supported by Taobao multimodal searchImage/text/audio/video
Duration of the MLEIL2026 conference3 days

Sources

  • Sohu Tech: iFlytek Spark Voice Foundation Model SparkAudio 1.0 Released
  • The Korea Times: Nearly Half of South Korea's New AI Models Are Built on Chinese Foundation Models
  • Sina AI Hotspot Hourly Report: October 25, 2026
  • Toutiao: Daily AI Brief (2026.10.25)
  • Toutiao: Today's AI Industry Hotspots at a Glance (2026-10-25)
  • Sohu News: MLEIL2026 Opens in Changsha
  • CSDN: AI Industry Daily, October 25, 2026