TechHorn Lab @TechHornLab

Robots can hurt. Badly. A parent in China says her young daughter was kicked in the mouth by a robot, knocking out four teeth. She may have to live on liquid food for quite some time. People say fists and feet have no eyes. Robot fists and feet don’t either — and they hit a hell of a lot harder. #Robotics #HumanoidRobots #ChinaTech #AI

keyboard_arrow_left Previous keyboard_arrow_right Next

Chinese EV makers are studying AI model architectures just as seriously as traditional AI labs and robotics companies. From Mixture-of-Experts-style architectures and on-device compute allocation to world models, VLA systems and the broader idea of “physical AI,” these companies are no longer merely figuring out how to put AI inside a car. A vehicle is already one of the most complex, data-rich and commercially viable pieces of AI hardware in existence. Xiaomi now has its own MiMo foundation model. It would hardly be surprising if XPeng, Li Auto and other Chinese automakers eventually released their own general-purpose models—or even open-sourced them. Large models are becoming the operating systems of intelligent vehicles. For Chinese EV makers, building their own may eventually become unavoidable. \#ChinaAuto #AI #Xiaomi #XPeng #LiAuto

XPeng is trying something remarkably bold: before its driver-assistance system acts, it first plays out the future inside its “head.” The upgraded second-generation VLA introduces X-Foresight, a predictive world model that uses the positions, speeds and movement patterns of nearby cars and pedestrians to simulate what could happen up to six seconds ahead. It does not predict just one outcome. Using Flow Matching, the model can generate multiple possible futures—a lead car suddenly braking, a pedestrian changing direction or another vehicle cutting in—then choose a more appropriate driving trajectory. Meanwhile, Infini-VLA remembers up to 30 seconds of previous road information. Streaming autoregressive inference processes the constantly changing present, while a Mixture-of-Transformers architecture dynamically assigns urban driving, campus navigation and parking tasks to different “experts.” Remember the past, understand the present and anticipate the future. XPeng also claims the upgraded model responds 300% faster end to end, with the G9L becoming the first production vehicle to receive it. The system cannot literally see the future. It imagines several plausible dangers before deciding what to do next. At least among production cars, that is a path we rarely hear automakers discussing so openly. \#XPeng #VLA #AutonomousDriving #ChinaAuto

keyboard_arrow_left Previous keyboard_arrow_right Next

At Gamescom, CD Projekt Red unveiled The Witcher 3 Remastered and a new expansion, Square Enix showed Final Fantasy VII Revelation, and Xbox released new footage of Gears of War: E-Day. Ubisoft’s most talked-about “new release,” however, was only 23.49KB. Ubisoft suddenly released the original Heroes of Might and Magic III Complete Edition on Steam for $9.99. After paying and downloading it, players found only a placeholder file named Heroes3.exe. When opened, it helpfully displayed: “Dummy executable.” Basically, Ubisoft uploaded the equivalent of a desktop shortcut as the game itself—and charged players $9.99 for an almost empty folder. #Ubisoft #Gamescom2026 #Steam #Heroes3

China is moving remarkably fast—and aggressively—on autonomous-driving legislation. The latest proposed amendment to China’s Road Traffic Safety Law says that when an autonomous-driving system is active, the vehicle’s manufacturer or importer would be responsible for handling traffic violations. If the company claims the violation was unrelated to the system, it would have to prove it. Cars equipped only with driver-assistance features would still be treated like conventional vehicles. Put that next to Chinese automakers’ current obsession with “little blue lights,” and things get interesting. The lights were not invented by Chinese automakers, but in China they have evolved from simple status indicators into symbols of being “futuristic,” “high-tech” and “ahead of the competition.” Automakers have made intelligent driving a major selling point while carefully keeping its legal classification at Level 2. Once a vehicle genuinely enters Level 3, control begins shifting to the system—and responsibility begins shifting toward the automaker. That is why the market has produced labels such as L2+, L2++, and the jokingly named “L2.99999.” Now the draft law proposes drawing the line according to whether autonomous driving is actually active. It may become harder for automakers to market systems as nearly Level 3 while leaving all the risk with the driver. Has China finally punctured the paper-thin barrier between L2.99999 and Level 3? #AutonomousDriving #ADAS #L3 #ChinaAuto

I’ve been having a great time watching the World Humanoid Robot Games lately. Honestly, it feels more entertaining than watching human athletes. Even weightlifting—which I usually find painfully boring at the Olympics—suddenly becomes fascinating when robots do it. #HumanoidRobots #RobotGames #Robotics #China

Yesterday, Xiaomi unveiled three new Xring chips: Xring O3, a flagship 3nm smartphone SoC with a claimed AnTuTu score above 5.22 million; Xring O100, an edge-AI chip offering 1.22TB/s of bandwidth and up to 330 tokens per second; Xring D100, an autonomous-driving chip supporting 160GB of unified memory and, according to Xiaomi, local deployment of 200B-class models. But the launch also produced an interesting argument. Some people say these chips cannot represent real technological breakthroughs, no matter how impressive their specifications look. Their reasoning: in China, any company that achieves a genuine breakthrough would already have been sanctioned by the US. Does that argument make any sense? #Xiaomi #Xring #Semiconductors #ChinaTech

Another fun video: China’s soccer robots have already mastered the true essence of the game: Diving. #RobotSoccer #EmbodiedAI #ChinaTech

In China, robots have become so widespread that they’re even being used for begging. #Robotics #EmbodiedAI #ChinaTech

In China, you can even watch a robot throw a tantrum. #WRC2026 #Robotics #ChinaTech

Move away from cloud subscriptions and onto personal computers, and local AI is increasingly starting to look like China’s game. For LLMs, there is Qwen3.8-27B. Quantized to Q4, it can deliver remarkably strong performance and speed on a single 16GB GPU. For video, there is the open-weight MiniMax H3, which understands text, images, video and audio—and generates video with native sound. MiniMax has also just launched MiniMax Design, a multi-agent workspace that brings scripts, images, video, voiceovers and editing into one production flow. We are not yet at the point where one high-end PC can run the entire process fully offline. H3’s complete 2K workflow still relies on APIs, while MiniMax Design does not run every underlying model locally. But for anyone with a powerful computer and an interest in local AI, this is unquestionably one of the most exciting times ever. And increasingly, the models and tools worth experimenting with are coming from China. #Qwen #MiniMax #LocalAI #OpenModels

A little over a year ago, I ran DeepSeek-R1-Distill-Qwen-32B locally on a PC equipped with an RTX 4080. It was a 32-billion-parameter distilled model with a roughly 20GB Q4-quantized file. Since it could not fit entirely inside the 4080’s 16GB of VRAM, it generated only about 5 tokens per second—basically spitting out the answer one word at a time. Yesterday, I ran the newly open-sourced Qwen3.8-27B on the same machine. This 27-billion-parameter dense model, also quantized to Q4, briefly reached 26 tokens per second. It wasn’t a strictly controlled performance comparison. But here’s the point: The same graphics card that could barely run a local LLM a little over a year ago can now generate answers at close to normal reading speed. AI models are evolving unbelievably fast. #Qwen #DeepSeek #LocalAI #RTX4080

n the US, Tesla uses its own Grok. In China, it has replaced it with ByteDance’s Doubao. Following software update 2026.14.13, Doubao now appears as a standalone app in Chinese Tesla vehicles. It offers four voices and can answer real-time questions, hold conversations, sing, debate, tell stories and help users practice English. It costs just RMB 9.99, or about $1.50, per month through Tesla’s Premium Connectivity service. But compared with the full US version of Grok, one crucial capability is still missing: Doubao currently cannot control the car. Grok can already set navigation, make calls, search for and play music, adjust cabin temperature and even open the glovebox. It also offers personalities ranging from Storyteller to Unhinged. In short, Doubao can keep Chinese Tesla owners company. Grok can actually get things done. #Tesla #Doubao #Grok #ByteDance

China’s World Robot Conference is now underway in Beijing, with more than 300 companies showcasing over 2,000 robotics products to packed crowds. Outside the venue, Unitree—freshly listed in Shanghai yesterday—is plunging. Unitree went public at RMB 150.80 ($21), opened yesterday at RMB 1,100 ($153), and closed at RMB 845 ($118). This morning, it fell as low as RMB 700 ($97), down more than 17% from yesterday’s close and 36% from its opening price. #Unitree #WRC2026 #Robotics #ChinaStocks

Unitree just unveiled a robot that it claims can run faster than Usain Bolt. Today, its stock followed suit. The Chinese robotics company made its A-share debut at an IPO price of RMB 150.80, or about $21 per share. It opened at RMB 1,100—up 629.44%—briefly giving Unitree a market value of RMB 444.9 billion, or roughly $62 billion. Investors who won one IPO lot paid RMB 75,400 for 500 shares. At the opening price, their paper profit reached RMB 474,600—about $66,000. The stock quickly pulled back toward RMB 900, but was still trading nearly 500% above its IPO price. Unitree’s robot just learned to jump two meters. Its stock has already jumped into the sky. #Unitree #Robotics #ChinaTech #IPO

Xiaomi is starting to treat itself as an AI company, too. The company reported RMB 108.92 billion—about $15.2 billion—in Q2 revenue. Management said API calls and Token Plan subscriptions for its MiMo model have begun generating revenue, with sales growing rapidly. Xiaomi also spent RMB 9.23 billion, or roughly $1.29 billion, on R&D during the quarter, up 18.9% year over year. Capital expenditure for its EV and AI businesses reached another $332 million. MiMo has actually received solid feedback, while usage and its user base continue to grow. It is no longer just an internal technology project—it has started making money. Selling phones and cars has become increasingly difficult work. For a company spanning smartphones, EVs, IoT, and robotics, AI may become Xiaomi’s most important new battlefield. #Xiaomi #MiMo #ChinaAI #AI

An RTX 5090 D V2 worth roughly $2,800 on China’s used market was offered just $1,880 by Zhuanzhuan, a Chinese electronics buyback platform. At that price, even Jensen Huang might want to buy his GPUs from them. Even more absurdly, used RTX 4090 cards are listed on Zhuanzhuan for around $2,770. In other words, its offer for the newer 5090 was nearly $900 lower than the price it charges for the older 4090. In another video from the same series, an RTX 5080 received an offer of just RMB 6,000—about $830. The Zhuanzhuan-affiliated creator explained that the video was recorded in July, during a slow period for GPU sales, and claimed that RMB 13,500 was a reasonable offer at the time. But no matter how slow the market was, an RTX 5090 was never selling for that price. In fact, this card has barely had an “off-season” since it launched. #Zhuanzhuan #RTX5090 #NVIDIA #GPU

Dreame may be in serious trouble. Xiaohongshu（Red）, China’s Instagram-like lifestyle platform, has suddenly filled with people publicly demanding payment from the company. Here’s what the screenshot says: Dreame is accused of owing supplier invoices, labor costs for its AWE exhibition, legal fees, model fees, franchise deposits, store renovation bills, and wages owed to former employees. One contractor is chasing a RMB 19,500 video-production payment. A potential franchisee says a RMB 50,000 deposit has not been returned. Creditor chat groups reportedly now have more than 700 members. Dreame denies that its major-appliance business has been dissolved and says some disputed payments are being processed. It is still too early to say Dreame has collapsed. But Xiaohongshu is starting to look less like a lifestyle app and more like Dreame’s online creditors’ meeting. #Dreame #ChinaTech #SmartHome

I saw a Chinese media report this morning saying DeepSeek’s official API price hike reached as high as 1,100%. We covered the planned increase last week and pulled together two sets of key comparisons. Some rates rising 500% already felt a little absurd. Now it’s 1,100%... But the math actually checks out. V4 Pro cached input rose from RMB 0.025 to RMB 0.30 per million tokens during peak hours. The new price is 12 times the old price, which equals a 1,100% increase. Even more counterintuitively, DeepSeek—the famous “price killer”—is now more expensive than GPT‑5.6 Luna on parts of its API. V4 Flash uncached input costs more both off-peak and at peak. Its output also costs more during peak hours. Only cached input remains cheaper. #DeepSeek #OpenAI #ChinaAI

@GrantAbrahamA @AiTechHubs Be careful！