Discover the AI Tools

A new leak suggests Moonshot's next flagship model, Kimi K3.1, could arrive as soon as next month with three reasoning intensity settings and a million-token context window. The details, pulled from internal API responses, also point to Agent mode and Swarm multi-agent collaboration. Moonshot hasn't confirmed anything yet, but the clues are adding up.

训练大模型强化学习，GPU烧钱如流水？腾讯混元团队从经典临界批量理论中找到突破口，重新定义在线RL的批量大小。结果很实在：PPO生成吞吐最高提升2.29倍，GRPO省下29%训练时间。原来批量不是越大越好，找准甜点区才是关键。

Alibaba DAMO Academy just unveiled DAMO EAGLE, an AI that catches esophageal cancer early using plain CT scans—no endoscopy, no contrast agents. Validated on 80,000 cases across three countries, it hit 90% sensitivity and 99.2% specificity. The goal? Flag high-risk patients so they can get endoscopy sooner, when cure rates top 95%. The paper landed in Nature Medicine on September 22.

China's live-streaming e-commerce is nearing 7 trillion yuan in annual sales, but as AI anchors cut costs by up to 90%, industry experts and top hosts like Li Jiaqi insist that real human connection remains irreplaceable. The future likely blends AI efficiency with human trust.

Huawei just launched Xiaoyi Work, an AI assistant that handles market research, PPTs, data analysis, and even coding—all from a simple sentence. It runs on HarmonyOS devices and can work in the cloud while you're away. New users get 1000 free AI points for a month. Could this be the office game-changer we've been waiting for?

Zhiyuan just hit a huge milestone: its 20,000th Yuzheng A3 Ultra robot rolled off the line, and 300 bots are now working daily at Chimelong's theme park. In just six months, production doubled from 10,000 to 20,000 units. This isn't a demo anymore—it's real work in crowded, unpredictable spaces.

腾讯推出了一款叫Hy Translate的翻译应用，支持33种语言双向翻译，还能离线使用。它瞄准的是出境游和商务沟通场景，目前已在美国、日本、韩国和东南亚等地上线。除了主流语言，它还覆盖藏语、维吾尔语、蒙古语、哈萨克语和朝鲜语。语音、拍照、对话、离线翻译一应俱全，没网也不怕。

At Meta Connect, Zuckerberg showed off a lineup of AI glasses and a new lightweight VR headset. The star? Muse, an AI assistant that's baked into the specs, has its own email, and might just become your new best friend. But will it avoid the fate of other AI gadgets?

OpenAI just gave ChatGPT a voice that actually does things. Pro and Plus users can now speak to create documents, emails, and even websites. Free users get app integrations too. It's a shift from answering questions to completing tasks—all without touching a keyboard.

Moonshot AI's next flagship, Kimi K3.1, may launch as early as October, according to leaked internal files. The model reportedly offers three reasoning levels—Low, High, and Max—plus a million-token context window and a Swarm mode for multi-agent collaboration. While nothing is official yet, the code hints at a major leap from single-model chat to team-based AI workflows.

Imagine asking your glasses what you're looking at—no phone, no Wi-Fi. At the 2026 Snapdragon Summit, Qualcomm and PrismML showed a 2-billion-parameter vision model squeezed into 1-bit precision. It runs entirely on the Snapdragon AR1 Gen 1 platform, cutting memory to a quarter of 4-bit models while doubling token speed. That means real-time object ID, translation, and voice help without the cloud—or the battery drain.

Google DeepMind's Demis Hassabis just confirmed that Gemini 4 is in post-training and could launch well before the end of the year. The company is skipping Gemini 3.5 Pro to focus on this next flagship, already testing it in internal coding tools and safety checks. Hassabis says he's '100% confident' in the team, and he's not worried about the AGI debate—he's focused on building trustworthy agents.

国产大模型领头羊DeepSeek年化收入已突破10亿美元，较几个月前翻倍。CEO梁文锋在投资者会议上透露，公司计划融资500亿元人民币，投后估值达5000亿元，并积极筹备上海证券交易所上市。这一迅猛增长主要得益于产品提价与模型采用率上升，为生成式AI商业化注入强心针。

Huawei just dropped Xiao Yi Work, an AI assistant for office tasks, coding, and creative projects. It syncs seamlessly across HarmonyOS phones, tablets, and computers. For a limited time, new users get 31 days of free trial with 1000 AI points. Curious how it works? It breaks down tasks, runs them on cloud or local machines, and keeps you in the loop.

OpenAI, Anthropic, and Hugging Face want countries to adopt shared safety tests and incident reporting for AI. But the Trump administration says that's a step too far, rejecting any global governance with real power. The fight isn't about whether risks exist—it's about who gets to set the rules.

Tencent just launched a new translation app that speaks your language—literally. The Tencent Hy Translation app handles 33 languages, including five minority dialects, and works offline. With voice, photo, and text translation, it's now available in 12 countries and regions across Asia. Built on the open-sourced Hy-MT2 model, it promises smoother, more accurate translations than ever before.

Ant's InclusionAI just dropped a transparency bombshell on HuggingFace. The AI-Transparency repo reveals what data trained Ling, Ring, and Ming models, all to comply with the EU AI Act. No weights, just documents. It's a rare peek behind the curtain of AI development.

China's first AI-generated long drama, *The Later Journey to the West*, just aired on Hunan TV. No cameras, no sets—every frame came from the Seedance model. A 100-person team restructured the entire filmmaking process, slashing per-episode costs from millions to tens of thousands. It's a quiet revolution in how TV gets made.

Ever notice Claude's writing feels off? An Anthropic engineer spills the beans: the model was optimized for math and code, so it learned to write for other AIs—dense, jargon-heavy, and hard for humans to digest. The good news? The latest version is finding its voice again.

Google just dropped two text-to-speech models that let you craft custom voices with simple descriptions—no technical know-how required. The Gemini 3.8 Flash TTS even replicates a voice from just 30 seconds of audio. With support for over 100 languages and 2,000 ready-made voices, this could change how creators bring characters to life.