Tencent’s New AI "T1" Just Proved China Is the New AI Superpower

Tencent just dropped its powerful new AI model, T1, and it's already going head to head with Deepseek's R1, even outperforming it in some areas, while also taking on OpenAI's O1. With faster responses, smarter reasoning, and a bold pricing strategy, Tencent is making some serious moves in the AI race. Here's what's going on all right, let me get you up to speed on the latest news from Tencent. This Chinese tech giant, which you might know as the company behind WeChat and massive video game hits like Honor of Kings, just officially launched a new AI reasoning model called Hunyun T1, often shortened to T1. Some people also call it Hyunyun Thinker 1. Now, there was a T1 preview floating around earlier that you could try out in Tencent's Yuanbao Chatbot, but the official version just arrived on the scene, and it offers faster responses, improved clarity, and some impressive performance numbers on various benchmarks. So let's dig into those performance details. Tencent said T1 scored 87.2 on MMLU Pro test. Deepseek's competing model R1 scored 84 on that same test. Meanwhile, OpenAI's 01 scored 89.3. That sort of puts T1 right in between the two, which is not bad at all, especially if you consider that T1 was just officially launched. Tencent is really emphasizing T1's ability to handle long text processing and produce results with minimal hallucination. They also note T1's clarity and neatness in text output, which is always helpful if you're dealing with super long docs or complicated queries. Now, if you're into math or logic, you might want to pay attention to T1's scores on the American Invitational Mathematics Examination, known as AIME, and a Chinese test suite called See Evil. On AIME 2024, T1 scored 78.2, which is just a bit behind R1's 79.8 and 01's 79.2. On the See Evil test, T1 actually tied with R1 at 91.8, outperforming O1, which got 87.8. So basically, T1 is showing strong results in both English and Chinese tasks. Now. One reason folks are excited about T1 is that it's built on a fancy architecture called Hybrid Transformer Mamba, or sometimes just Mamba. Tencent says that combining Google's Transformer approach with Mamba can cut down on memory usage in a pretty significant way. They're calling it the world's first large scale model to incorporate MAMBA in this manner. And what that means for users is that T1 can manage big context without demanding monstrous amounts of computing resources. In fact, they claim a 200% increase in decoding speed compared to traditional architectures, which could be a game changer if you're running AI workloads at a huge scale. They even said T1 invests something like 96.7% of its computing power in reinforcement learning, which is basically teaching the model through trial and error based feedback. That's also a pretty serious shift in training strategy. So how does pricing fit into all of this? DeepSeek's R1 model got a lot of attention, partly because it offered super aggressive pricing. Now Tencent's T1 tries to rival that, charging 1 yuan around 14 cents in US dollars per 1 million tokens for input and 4 yuan per 1 million tokens for output. Deepseek R1 charges 1 yuan per million tokens for input during the day and 0.25 yuan at night, while output is 16 yuan in the day, dropping to 4 yuan overnight. Depending on how you use these models, that cost difference can matter a lot, especially if you're feeding the model huge amounts of data or generating massive text outputs. Essentially, T1's cost structure is in the same neighborhood and might be even more appealing if you're only focusing on daytime usage, although R1 does offer those cheaper nighttime rates. Now, since we're on the topic of Tencent, you should know they've been doubling down on AI investments like crazy. They poured $10.7 billion into capital expenditures in 2024, up from 3.4 billion the year before. In the fourth quarter alone, they spent another $5.4 billion on AI initiatives. Capital expenditures are now around 12% of their total revenue, and if that sounds impressive, it's because it is. They're not shy about telling the world how serious they are about AI. They've mentioned reorganizing their AI teams, boosting research and stepping up marketing for ainaps products. And obviously they plan to ramp up capital spending even more in 2025. Tencent's not the only one going big on AI in China. Alibaba recently said they'd invest at least $452.6 billion in cloud computing and AI over the next three years. And ByteDance, the company behind TikTok, set aside over $20.7 billion for 2025, much of which is also for AI infrastructure. Then you have Deepseek, the startup that came out with R1, which has been a real disruptor. Their R1 model even rattled the global equities market not too long ago, causing a sell off worth around a trillion dollars when it was first unveiled. Now it's going toe to toe with OpenAI's advanced models and apparently Tencent is making sure they stay in the race. Let's circle back to T1 for a second. Tencent tested it on various internal and external datasets including MMLU Pro code evaluation benchmarks like Live code bench, where T1 scored 64.9 and advanced math tests like Math 500 where it got 96.2, really close to R1's performance. We're also hearing about something called GPQA diamond, dealing with high level questions in physics, chemistry and biology, and T1 scored 69.3 there. Meanwhile, in different alignment tasks that measure how well the model follows human instructions, T1 did well with a 91.9 in something called Arena Hard. These are fairly specialized metrics, but they matter because they show how robust the model is in real world tasks like like coding, scientific problem solving and obeying or interpreting user requests. Speaking of user requests, T1 is integrated into Tencent's Yuanbao chatbot, but here's something interesting. Tencent also uses DeepSeek's R1 in Yuanbao. In fact, Tencent's chairman and CEO Pony Ma, also known as Ma Huateng, has publicly said he admires Deepseek's open source approach and is adopting a double query strategy, giving users access to both T1 and R1. This is pretty unusual since you often see big companies focusing only on their in house solutions, but Tencent is basically saying we want the best of both worlds. That flexibility may be part of how they plan to keep WeChat, which has over a billion users at the forefront of AI experiences. Plus, if you're a developer or business using Tencent Cloud, you can apparently choose between these AI models for different tasks or maybe combine them, which is quite fascinating. In terms of financials. Beyond their capital expenditure, Tencent reported revenue of 172.4 billion yuan, about $23.9 billion in the fourth quarter of 2024, beating the forecast of 168.9 billion yuan. Their profit hit 51.3 billion yuan, way above the expected 46 billion. So that's an 11% year over year revenue jump and a 90% hike in profit. Gaming was a massive contributor. Domestically, gaming revenue rose 23% to 33.2 billion yuan and internationally it climbed 15% to 16 billion. Advertising revenue also grew 17% to 35 billion, partly because Tencent says its AI powered advertising infrastructure can deliver more relevant ads. Their AI cloud revenue, by the way, about doubled in 2024, though we don't have the exact figure, which also shows how crucial AI is becoming to their business model. Those higher revenues, however, came alongside a big spike spending on GPUs and servers critical for running these monstrous AI models. Tencents explained that some of these GPUs are allocated for internal use, which is limiting how many they can offer to external cloud clients at the moment. But they expect that as they step up GPU purchases, the cloud services revenue could accelerate. It's all part of this massive AI race in China, where big names like Tencent, Alibaba, Baidu and ByteDance are moving quickly to cement leadership in generative AI, large language models and advanced reasoning systems. And hey, if you're wondering about the man at the helm, Ma Hua Tung, also known as Pony Ma, has a real time net worth of about $54.1 billion, according to Forbes. That's as of March 23, 2025. Not too shabby, right? He's definitely banking on AI to keep those numbers growing. The entire market in China is somewhat spurred on by the success of success of Deepseek, which has proven that homegrown Chinese tech can rival western counterparts like OpenAI. Tencent has even said it's impressed with how Deepseek has provided a truly open source approach and we're seeing this two pronged approach using both Tencent's own Hunyuan T1 and Deepseek's R1, mirroring what Tencent did in gaming, where they invest in both in house and third party games. On top of that, companies like Alibaba have launched their own AI models such as qen, while Baidu just released fresh reasoning models of its own. Everybody's trying to differentiate, whether by offering faster speeds, better language coverage or more affordable pricing. T1's approach with that Mamba mo e architecture is all about balancing computational efficiency with advanced reasoning. Meanwhile, R1 took the world by storm, partly because it arrived earlier and and boasted both strong performance and super competitive pricing. That's why T1's cost structure is so important. Tencent can't just rely on name recognition if it wants to compete with a popular startup. Essentially, Tencent's T1 model combines low costs with powerful reinforcement learning, using a curriculum learning approach and a self rewarding system to improve problem solving and text generation. With the company heavily investing in AI. T1 is emerging as a strong competitor to Deepsea seeks R1, showing strengths in logic science and long text processing, with potential expansion into WeChat features and enterprise analytics. Alright, that's going to do it for today's rundown. If you like this breakdown, don't forget to drop a comment and let me know what you think. Think these huge investments in AI are going to pay off. I love hearing your thoughts. Thanks for watching and I'll catch you in the next one. It.