Baidu Wenxin 5.0 and Tesla’s AI5 chip debut together! → The latest AI news is here!

AI Frontier Express ↓↓

01 Baidu’s Dual-Engine Upgrade:

The Wenxin Large Model 5.0 + GenFlow 3.0 empowers everyone to become a "super individual."

On November 13, 2025, at the Baidu World Conference, Wenxin Large Model 5.0 and Wenku Online Storage GenFlow 3.0 were officially launched. Wenxin 5.0 is the world’s first natively multimodal large model, boasting a parameter count of 2.4 trillion. It supports unified understanding and generation across multiple modalities, including text, images, audio, and video. Also unveiled simultaneously, the GenFlow 3.0 agent platform has over 20 million active users and introduces two new agents—Office Agent and GenX Agent—that help users work more efficiently and unleash their creative potential, fully empowering the “super individual.”

02 Byte TRAE SOLO Official Version Now Available—Limited-Time Free Trial for International Users!

On November 12, TRAE, ByteDance’s AI programming tool, officially launched the full version of SOLO, positioned as “The Responsive Coding Agent.” International users can enjoy a limited-time free trial from 12:00 on November 12 to 23:59 on November 15. The full version introduces two new agents—SOLO Coder and SOLO Builder—that offer core capabilities such as multi-task parallelism, context compression, and code change tracking, enabling users to complete the entire development process—from requirement analysis to deployment—in one integrated platform.

03 A 1.5-billion-parameter small model outperforms a 671-billion-parameter giant model, with training costs under $8,000.

On November 13, 2025, Weibo officially open-sourced its first self-developed large language model, VibeThinker-1.5B. Despite having only 1.5 billion parameters, this model outperformed the DeepSeek-R1 model—whose parameter count is over 400 times greater (671 billion)—in mathematical competition tests such as AIME24/25 and HMMT25. It also surpassed commercial models in the LiveCodeBench V6 code benchmark. The training cost for this model was less than $8,000. By introducing an innovative “Spectrum-to-Signal Principle” (SSP) approach, it has challenged the industry’s conventional wisdom that “intelligence is determined by parameters.”

04 Microsoft Launches “Planetary-Scale AI Superfactory”:

The distributed computing power revolution has shortened training time by 70%.

On November 12, 2025, Microsoft will officially launch its first “Planet-scale AI Superfactory,” connecting its two Fairwater data centers located in Wisconsin and Atlanta via a dedicated AI WAN network to create a virtual supercomputer. This system integrates hundreds of thousands of GPUs, EB-level storage, and millions of CPU cores, enabling AI model training tasks that would normally take months to be completed in just weeks. At the same time, Microsoft announced plans to double the size of its data centers over the next two years, further strengthening its AI infrastructure and leading the global AI computing power race into a new era of distributed collaboration.

05 Apple Launches AI Breakthrough Battle:

A self-developed model with trillions of parameters aims to rival Google; Siri will first rely on its competitor’s “brain” as a transitional solution next year.

Apple is accelerating its push toward a self-developed AI strategy, planning to launch Apple Intelligence—a foundational model with a parameter scale of 1 trillion by 2026. The goal is clear: to reduce Apple’s reliance on Google’s technology. Currently, Apple has already partnered with Google, paying $1 billion annually for access to a customized version of the 1.2-trillion-parameter Gemini model to upgrade Siri. This model will take on intelligent summarization and task-planning functions for the voice assistant. Meanwhile, Apple’s in-house, self-developed model remains immature due to challenges such as talent attrition. Notably, Apple internally refers to the Google model as “AFM v10” to downplay the collaboration’s traces, while also deploying it via private cloud to ensure user data remains isolated from Google. This “competitive collaboration” underscores Apple’s strategic ambitions in the AI era: in the short term, leveraging Google’s technology to maintain Siri’s competitiveness; in the long term, using its own self-developed model to reshape its ecosystem’s power dynamics.

06 Musk Unleashes His Big Move!

Tesla to Build Trillion-Dollar Chip Factory; AI5 Chips, Which Outperform All Others, to Enter Pilot Production Next Year

At the 2025 shareholder meeting, Tesla CEO Musk announced the construction of a massive chip factory called "TeraFab," with a monthly production capacity of one million wafers. The initial target is 100,000 wafers, and the company will simultaneously advance the research and development and production of its fifth-generation AI chip, AI5. AI5 is fabricated using a 3nm process and boasts computing power ranging from 2,000 to 2,500 TOPS. Its power consumption is only one-third that of NVIDIA’s Blackwell chip, and its cost is less than 10% of the Blackwell chip’s price. AI5 is specifically designed for autonomous driving and the Optimus robot. The chip will be jointly manufactured by TSMC and Samsung, with pilot production starting in 2026 and mass production scheduled for 2027. At the same time, Tesla is considering partnering with Intel to ease supply-chain pressures. This move marks a critical step toward Tesla’s vertical integration of the chip supply chain and is expected to reshape the global semiconductor industry landscape.

07 Meituan’s first AI-powered programming tool enters public beta:

Agent-driven code generation boosts programming efficiency by over 50%.

On November 10, 2025, Meituan officially launched the public beta of CatPaw, its first AI-integrated development environment (IDE) product. This tool is centered around agent-human collaboration and adopts an intelligence-driven programming paradigm, offering features such as code completion and project preview/debugging. It is powered by Meituan’s self-developed LongCat model and supports mixed calls across multiple models. Its Ask mode is ideal for simple question-and-answer scenarios, while the Agent mode can automatically analyze the project context to enable code block prediction, batch modifications, and one-click fixes for frontend errors. Internal data shows that the tool has achieved a R&D penetration rate exceeding 95% at Meituan, with an AI-generated incremental code rate surpassing 50%. During the public beta phase, the tool supports Mac systems, and new users who sign up will receive 500 free conversation credits. A Windows version is coming soon.

Editor: Yu Li

Proofreading: Zeng Xiong

Reviewed by: Zhang Weitian