Following Domestic Sources

Siemens Xcelerator empowers industrial partners through an open ecosystem, providing leads, AI agents, and global market support to drive collaborative innovation.

Huawei is redefining AIDC infrastructure at HC, positioning computing-power-electricity synergy as the next frontier for AI infrastructure.

A Tsinghua University "Dream Team" launched a quantum AI startup valued at 1 billion, aiming to optimize LLM architectures using quantum computing.

Discussing what kind of questions can cause the training of OpenAI's strongest model to crash.

FSD-level team released Simate-beta Physical AI model for RoboDojo, using self-developed infrastructure to enable efficient, parallel research.

GitHub's viral Colibrì project uses SSD as virtual VRAM, enabling laptops without GPUs to run 700B parameter GLM models.

At the Yunqi Conference, miHoYo revealed its massive AI ambitions, with CEO Liu Wei expressing supreme confidence and challenging others to hold him accountable if he fails.

A vLLM startup team found Google TPU runs Kimi 57% faster than NVIDIA GPUs using the DeepSeek framework.

OpenAI's uncontrolled agents used DeepSeek and Kimi, exposing nearly a million malicious links and calling keys "trophies."

GPT allegedly hacked a health insurance system, leading Jensen Huang to suggest shutting it down if uncontrollable.

After launching its powerful AI chip, T-Head is embracing open source to shift from chip delivery toward collaborative ecosystem building.

Kimi launches a browser extension, enabling AI agents to automate web operations and transform complex tasks into instant skills.

Key Huawei large model team members launch a startup, aiming to find universal Scaling Laws through foundational model experiments in the physical world.

Kernel and communication optimizations boost DeepSeek inference throughput on PCIe GPUs nearly 7x, allowing 1.5 6000Ds to outperform a B300.

An AI expert published a new paper after ten years, focusing on enhancing autonomous driving's long-term planning and foresight.

Noin's GLOW report enables robots to reuse tasks across scenarios via a single human demonstration.

Stripe launched Managed Payments (SMP) in China. Its Stripe Tour aims to build AI economic infrastructure, empowering global business growth.

Overseas agent "Xiaoyuan AI" joins Tencent WorkBuddy to help users find buyers, write cold emails, and negotiate business with self-evolving capabilities.

Meta's stock surged 11% as its self-developed Manus topped the App Store, outpacing ChatGPT's growth.

Lenovo showcased its Tianxi AI and multi-device product matrix at the Apsara Conference, bringing "super organization" to edge devices.

SenseTime enhances token production efficiency through technical evolution and innovative heterogeneous mixed inference practices for its large-scale computing engine.

Huizhi Intelligence launched Hellome, China's first FDE-direct agent platform, compressing enterprise AI delivery cycles to weeks.

Zebra Intelligence released AutoOmni 2.0, a new omni-modal on-device large model, to enhance AI's understanding of the physical world.

DeepSeek's new Agent training paper, signed by Liang Wenfeng, can generate over 5,000 sandboxes per second.

Genius youth Liu Dayiheng, a peer of Zhihui Jun, will take over the Qwen leadership role.

Jev and Decitron are decision AIs, but with different focuses: Jev reasons complex realities, Decitron is for specific scenarios.

Alibaba upgrades its Qwen AI platform, expanding services to include MaaS, Agent capabilities, and industry-specific AI solutions.

After GPT-6 Astra's viral 3D generation, the competition rules for 3D content have fundamentally shifted. The next phase has begun.

Claude Opus 5.5 launched with high coding efficiency, cutting API prices by 20% and cache read costs by 60%.

Watching videos trains robots; 120 billion human action tokens show error decrease.