Ep 228: A Chinese company just released a massive new AI agent model trained entirely without any…

From Models & Agents for Beginners — Video Edition · 06:44

Show by Patrick

Models & Agents for Beginners A Chinese company just released a massive new AI agent model trained entirely without any Nvidia chips. Today we look at a giant new AI model built for agents that runs on Chinese hardware instead of the usual chips. This matters because it shows AI development is spreading beyond a few big players. We'll also break down how these agent models actually work under the hood and check out a shopping plugin that brings AI straight into a department store app. Plus a quick look at a free tool someone built to fact-check politicians live. The Big Story A big Chinese company called Meituan just released the weights for something called LongCat-2.5. It is a 1.6-trillion-parameter model made specifically to power AI agents, and the whole thing was trained on Chinese hardware with zero Nvidia GPUs involved. Think of it like building a super-smart helper that can plan and carry out tasks, except this one was made in a completely separate tech ecosystem. The model scores 88.4 on a test called Terminal Bench, which puts it just behind one of the top models from another lab. The sample deployment runs on 194 Huawei Ascend chips, and Meituan makes no mention of Nvidia GPUs anywhere in the announcement. What makes this exciting is that it proves powerful AI agents can be created without relying on the same small set of chip suppliers everyone else uses. For you, this could mean more choices in the future when you want an AI helper for school projects or creative work, because competition often leads to better and cheaper tools. It also hints that AI development is becoming more global, so the assistants you use in a few years might come from places you did not expect. Right now the weights are public, so researchers and developers can download and experiment with them, though a regular beginner would need a parent or teacher to help set it up on compatible hardware. The sample version runs on 194 Huawei Ascend chips, showing the whole system works end-to-end on non-Nvidia equipment. This release lands the model just behind Claude Opus 5 on the Terminal Bench score of 88.4 points. Because the training happened entirely on Chinese hardware, it demonstrates that large-scale agent models no longer need to depend on a single dominant chip maker. Beginners interested in how AI agents are built can watch for future demos that use these open weights once easier interfaces appear. Source: the-decoder.com Explain Like I'm 14 You know how when you text a friend, your phone sometimes guesses the next word before you finish typing? Now picture that same idea stretched across an entire team of helpers instead of just one word. One helper might look at your request and decide what steps are needed, another might search for information, and a third might check if the answer makes sense. Each of those helpers is like a small piece of the bigger AI, and the whole group works together without you having to tell every single step. The model from today is built exactly for this kind of teamwork, so it can handle longer jobs like planning a project or running repeated tasks in the background. What feels magical is that the AI is not just answering one question and stopping; it is keeping track of the goal and bringing in more help when the job grows. That is why people call these systems agents: they act more like teammates than simple chatbots. Imagine you ask the AI to organize a group project for school. The first part of the model breaks the project into deadlines, the second part looks up reliable sources online, and the third part writes a short summary you can review. All of this happens in one conversation thread instead of you sending separate messages for each piece. The 1.6-trillion-parameter size gives the model enough capacity to keep many of these helper roles active at once without losing track of the original request. Because the model was trained on Chinese hardware, the same teamwork approach can now run on different chip types, which may lead to more options for how these agent teams are hosted in the future. Cool Stuff & Try This AI that shops with you at a real department store Hyundai Department Store added a new plugin so ChatGPT can help you browse and buy things straight from their stores. Instead of opening a separate app and searching by yourself, you can ask ChatGPT questions about products, sizes, or recommendations and it pulls the info from the store’s own system. This feels useful if you like shopping for clothes or gifts but want quick suggestions without clicking through dozens of pages. The plugin extends AI shopping directly into the department store’s catalog, so answers come from real inventory rather than general web results. Anyone with a ChatGPT account can try the plugin once it rolls out. Go to the ChatGPT website or app, start a new chat, and type something like “Show me winter jackets from Hyundai Department Store under $100” to see how the plugin responds with real…

More episodes of Models & Agents for Beginners — Video Edition

More episodes

PodSearch is a free, independent podcast search engine. Read about PodSearch, the privacy policy or contact us.