24-hour AI news: Llama added to government list, domestically produced "trusted innovation module box" debuts, and large computing power orders continue to rise.

In the past 24 hours (September 22nd), Meituan released a new large-scale model branch and an industry foundation, and Beijing launched two AI events. Meanwhile, Llama received approval from US government agencies, NVIDIA established a joint laboratory in the Middle East, and OpenAI announced new high-computing features and a massive GPU purchase across the industry chain, echoing global developments. 1. Meituan Releases LongCat-Flash-Thinking, a High-Efficiency Inference Model. 1. Released on September 22nd and available in the open source community, it's positioned as "stronger reasoning, more professional." 2. Emphasizing the combination of "deep thinking + tool invocation" and "informal + formal" reasoning capabilities, it covers complex tasks such as mathematics, code, and agents. 3. Officially, it's the first open-source large-scale language model branch of its kind in China, continuing the "fast + powerful" approach of the LongCat-Flash series.

2. Fourth Paradigm Releases ModelHub XC, a "Intelligent Model Box."

1. Officially announced on September 22nd, it addresses the pain point of "computing power-model" incompatibility and connects domestic GPUs with domestic large models.
2. Introducing the EngineX adaptation engine, it adapted hundreds of models on the first day, with plans to expand to "thousands" within six months and "hundreds of thousands" within a year.
3. Currently covers domestic computing power such as Ascend, Cambricon, Tianshu Zhixin, Kunlun Core, Muxi, and Xiwang, focusing on "multi-model plug-and-play."

3. China Communications Information Technology Group Releases "CCCC Blue" Digital Base

1. Released on September 22nd, this is a complete set of autonomous and controllable digital bases for the transportation industry. 2. Establish ten centers: Blue Boat, Blue Hui, Blue Path, Blue Wing, Blue Alliance, Blue Star, Blue Ocean, Blue Stack, Blue Base, and Blue Shield, integrating AI, big data, IoT, Beidou, and security capabilities. 3. Focusing on serving scenarios such as smart highways, smart ports and shipping, and digital cities, it is considered a milestone in the "smart upgrade" of transportation infrastructure. 4. 2025 University Student AI Art Season Opens in Beijing 1. Launching on September 22nd and lasting for one month, the event will feature sections such as AI imaging, AI music, and AI interaction. 2. Nearly 1,600 works will be presented, covering over 160 universities. Courses and platforms will be co-developed with companies such as Keling AI, Jimeng AI, and Shengshu Vidu. 3. Simultaneously launch the "Intelligent Connectivity Community" and the "Global Call for Submissions" initiative to strengthen industry-university-research collaboration and international outreach. 5. The AI Creator Carnival concluded in Haidian. 1. The five-day event, themed "Let's in Vibe," concluded on the evening of September 21st. 2. Over 100 ecosystem partners, 40+ sub-venues, and 80+ interactive booths attracted over 10,000 participants. 3. The entire street was open for co-creation, focusing on practical scenarios such as "AI + Entertainment/Education," creating a "Borderless AI Practice Theater." 6. Meta Llama Receives Approval for Use by U.S. Government Agencies. 1. On September 22nd, the U.S. General Services Administration (GSA) added Llama to its list of available AI tools for federal agencies. 2. Governments can pilot open-source models under regulatory compliance, reducing pilot costs and procurement barriers. 3. Its inclusion on the list, alongside tools from AWS, Microsoft, Google, Anthropic, and OpenAI, provides significant endorsement of the open-source approach.

VII. NVIDIA and TII Abu Dhabi establish a joint AI and robotics lab.

1. On September 22, it was announced that the Middle East's first "NVIDIA AI Technology Center" has been established in the UAE.
2. NVIDIA's Thor chip will be deployed to advance the development of robotic platforms, including humanoid, quadruped, and robotic arms.
3. The lab targets applications in multiple fields, including climate, energy, and genomics, expanding regional computing power and talent pooling.

VIII. OpenAI previews "new high-computing features," some of which will only be available to Pro users.

1. On September 22, Sam Altman announced that several "computing-intensive" features will be launched in the coming weeks.
2. Due to cost constraints, some features will be available to Pro users first, and some may require additional fees.
3. The company's long-term goal remains "cost reduction, efficiency improvement, and expanded accessibility," with a focus on internal testing and tiered pricing in the short term. 9. MIT Launches SCIGEN: A "Constraint Generation" Tool for Quantum Materials. 1. Published in Nature Materials on September 22nd, this paper injects structural constraints (such as Kagome) into diffusion models. 2. It generates tens of millions of candidates, screens them, and synthesizes two novel compounds with unique magnetic properties. 3. It provides a new paradigm for the efficient discovery of materials required for quantum computing. 10. IREN Purchases Approximately $674 Million in GPUs, Aiming to Exceed $500 Million in ARR for its AI Cloud. 1. Announced on September 22nd: 12,400 new GPUs (approximately 7,100 B300s, 4,200 B200s, and 1,100 MI350Xs). 2. The total scale has increased to approximately 23,000 GPUs, with the goal of achieving annualized AI Cloud revenue exceeding $500 Million in the first quarter of 2026. 3. The Blackwell and MI series are being developed in parallel, emphasizing flexible provisioning and clustered delivery for training and inference. Frequently Asked Questions (Q&A) Q: What does Llama's inclusion on the US government's list mean? A: This means federal agencies can directly trial/procure large open source models under regulatory compliance, lowering the threshold for pilot testing and evaluation, and competing with multiple commercial closed-source solutions. This is expected to accelerate the deployment of low-cost AI in government processes (September 22). Q: Will OpenAI's "new high-computing features" be more expensive? A: In the short term, they will. Officials stated that some features are only available to Pro users, and some may incur additional charges. The rollout cadence is "coming weeks" because new capabilities require higher GPU computing power. Q: Why is IREN consolidating its procurement of the B300/B200/MI350X? A: To expand AI cloud provisioning and delivery capabilities for key customers. With the addition of 12,400 new licenses, the total number has reached approximately 23,000. The goal is to increase AI Cloud ARR to over $500 million by the first quarter of 2026, betting on the comprehensive performance of the first generation of Blackwell in training and inference.

Q: What problems do "CCCC Blue" and 4Paradigm ModelHub XC solve respectively?

A: "CCCC Blue" is the digital foundation for the transportation industry, solving the systematic implementation of cross-domain data and AI capabilities in industry scenarios; ModelHub XC is a "model-domestic computing power" adaptation platform, solving the compatibility and deployment efficiency in multi-model and heterogeneous domestic chip environments.