Home › News › AI Daily October 25 Noon Edition

AI Daily | October 25, 2026 Noon Edition: DeepSeek Open-Sources V4.1 and Opens Its Search API, European Banks Showcase NVIDIA AI Applications at GTC

Published: 2026-10-25 12:00  |  Category: AI Daily  |  Read: about 7 min

Top Stories This Noon

  • DeepSeek open-sources V4.1-Flash-1017: it also opens its search API, shifting the focus of large-model competition from "who is stronger" to "who is cheaper and who can deliver"
  • CAICT report: domestic token consumption averaged about 500 trillion per day in June: nearly four times the December 2025 level, as AI applications enter a phase of large-scale deployment
  • European financial institutions showcase NVIDIA AI applications at GTC Berlin: Revolut uses the PRAGMA model to serve 80 million customers, and Lloyds develops the ATUM trading model
  • The Ministry of Commerce proposes adding 40 technologies to the export control catalogue: AI and photovoltaics face systematic controls for the first time
  • Yizhu Technology completes a 1 billion+ yuan funding round: its compute-in-memory AI chip opens a new path, with a 3D DRAM architecture targeting edge inference bottlenecks

1. DeepSeek Open-Sources V4.1-Flash-1017 and Opens Its Search API

DeepSeekOpen SourceSearch API

On October 17, DeepSeek officially open-sourced the V4.1-Flash-1017 version and simultaneously opened its search API, marking an important step for Chinese large models in the open-source ecosystem and commercialization.

  • The V4.1-Flash version focuses on cost-effective inference and suits large-scale deployment scenarios
  • The open search API lets developers directly access DeepSeek's retrieval-augmented generation capabilities
  • The focus of large-model competition is shifting from "who is stronger" to "who is cheaper and who can deliver"
  • DeepSeek's funding is reported to have grown to around 100 billion yuan
Editor's note: DeepSeek has found a balance between open source and commercialization. Opening the search API not only expands application scenarios but also lays the groundwork for building a developer ecosystem. With the token price war intensifying, "cheap plus deployable" is becoming the core competitiveness of Chinese large models.

2. CAICT Report: Domestic Token Consumption Averaged About 500 Trillion Per Day in June

CAICTToken ConsumptionIndustry Report

The AI Research Institute of the China Academy of Information and Communications Technology (CAICT) jointly released the "Large Model Token Service Performance Monitoring Report (2026)", disclosing key data on domestic large-model usage.

  • In June 2026, domestic token consumption averaged about 500 trillion per day, nearly four times the December 2025 level
  • The normalized average price in China since 2026 has been 13.46 yuan, nearly double the 6.8 yuan in the second half of 2025
  • Overseas prices remain far higher than domestic ones, with the average normalized price of typical models at about 117.8 yuan since 2026
  • In August, the average TPS of domestic models reached 78.23 tokens per second, double the 2025 average of 38.25
  • The average domestic TTFT (time to first token) rose to 1.26 seconds
Data highlight: Average daily consumption of 500 trillion tokens is equivalent to about 350,000 tokens per person in China per day, reflecting how deeply AI applications have become integrated into daily life and production scenarios.

3. European Financial Institutions Showcase NVIDIA AI Applications at GTC Berlin

NVIDIAGTC BerlinFinancial AI

On October 25, 2026, at the GTC Berlin conference, several European banks, payment companies and financial data providers showcased AI applications built on NVIDIA accelerated computing, marking broad adoption of open-weight models in the financial industry.

  • Revolut uses NVIDIA's PRAGMA model to serve more than 80 million customers
  • Lloyds Bank developed the ATUM trading model to improve trading efficiency and risk control
  • SBS launched a sovereign intelligence platform to meet European data compliance requirements
  • These systems significantly reduce operating costs and drive the industry's intelligent transformation
Industry trend: The financial industry is moving from "trial-and-taste pilots" to "large-scale deployment." European banks' choice of NVIDIA open-weight models reflects both performance considerations and attention to data sovereignty and compliance.

4. The Ministry of Commerce Proposes Adding 40 Technologies to the Export Control Catalogue

Export ControlsPolicyAI Regulation

Recently, the Ministry of Commerce and the Ministry of Science and Technology issued Announcement No. 48 of 2026, seeking public comment on the revision of the Catalogue of Technologies Prohibited or Restricted from Export in China, proposing to add 40 technologies to export controls.

  • AI technologies and photovoltaic technologies face systematic controls for the first time
  • Covers core links such as large-model training methods, inference optimization and data processing
  • The move reflects heightened awareness of protecting key technologies amid intensifying international technology competition
  • Forms a reciprocal response to US and European AI export controls against China
Policy impact: The revision of the export control catalogue will profoundly affect the overseas cooperation models of domestic AI companies. Companies need to strike a balance between technology export and compliance while watching for follow-up implementation rules.

5. Yizhu Technology Completes a 1 Billion+ Yuan Round on a New Compute-in-Memory Chip Path

Yizhu TechnologyFundingAI Chips

The domestic compute-in-memory AI chip company Yizhu Technology has completed a 1 billion+ yuan funding round; its 3D DRAM architecture offers a new path for large-model edge inference.

  • Core technology: a compute-in-memory architecture that breaks the traditional von Neumann bottleneck
  • Adopts a 3D DRAM architecture to deliver higher compute density within limited power
  • Target scenario: large-model edge inference, reducing deployment costs
  • The funding will go toward volume production and market expansion

6. The National Data Group Is Established to Lower Enterprise Data Costs

National Data GroupData Elements

The National Data Group has been officially established to integrate public data resources and lower the cost for enterprises to obtain and use data.

  • The group will coordinate national-level data infrastructure
  • Promote reform of the market-based allocation of data elements
  • Provide more standardized, lower-cost data supply for AI large-model training

7. Geely Launches a 2,250 kW Smart Charging Robot

GeelySmart Charging RobotNew Energy

Geely has launched a 2,250 kW smart charging robot, achieving megawatt-level output from a single connector and charging from 10% to 70% in about four and a half minutes.

  • Charging upgrades from a "supporting facility" to an "intelligent connected terminal"
  • Megawatt-level charging power sets an industry record
  • AI scheduling algorithms optimize charging strategies and extend battery life

Today's Data at a Glance

MetricFigure
Domestic average daily token consumption (June 2026)About 500 trillion
Growth in consumptionNearly 4x (vs. December 2025)
Normalized average domestic token price (2026)13.46 yuan
Normalized average overseas model price (2026)117.8 yuan
Average domestic TPS (August)78.23 per second
Customers served by Revolut80 million+
Technologies proposed for export controls40
Yizhu Technology funding1 billion+ yuan
Geely smart charging robot power2,250 kW
10%→70% charging timeAbout 4.5 minutes

Sources

  • Toutiao: AI Highlights: DeepSeek Open-Sources V4.1, OpenAI Resumes Training, Capital Rushes In
  • Sohu Tech: CAICT: Domestic Token Consumption Averaged About 500 Trillion Per Day in June
  • Sohu Tech: European Financial Institutions Adopt NVIDIA Open-Weight Models to Advance AI Applications
  • Sina AI Hotspot Hourly Report: October 25, 2026
  • Toutiao: Tech Morning Brief, October 25, 2026
  • AI Agent Community: 2026 AI Frontiers: Latest Advances in Large Language Models