New Small Open Source AI Just Beat DeepSeek R1 and o1 mini Experts SHOCKED

So there's this division at Alibaba called the Quen team, and they've been cooking up a series of open source large language models. Their newest Release is called qwq 32B, which stands for Quen with Questions, and it's seriously catching everyone's attention because it claims to match or even beat Deepseek R1, a big time reasoning model that weighs in at a whopping 671 billion parameters. The real kicker is that the model only has 32 billion parameters, so on paper it's way smaller than deep seq R1. People are losing their minds about how a smaller model can pull off similar performance in math, coding and general reasoning tasks, especially when it usually requires way fewer resources to run. The Qin team says it can typically be deployed on hardware with only about 24 gigabytes of VRAM, which feels very light compared to more than 1,500 gigabytes of total VRAM usage. If you wanted to run the complete deepseek R1 configuration on multiple high end GPUs, that difference alone is making a lot of folks ask whether Small is officially the new Big in this AI arms race now. The QEN team first introduced this reasoning model back in November 2024 when they launched it as an open source rival to some of OpenAI's advanced reasoning models like O1 Pro. At the time, that first iteration already impressed people with its ability to do introspective step by step logic, especially for math and coding tasks. The trick, or at least one of the big ones, was that QVQ would pause mid answer to review what it was writing, but it had some issues on programming benchmarks such as Live Code Bench, and so some fans felt it still needed to catch up with OpenAI's coding models. The fascinating part was that Alibaba decided to release it under the Apache 2.0 license, which basically let anyone businesses, researchers or whoever tweak it, refine it, run it on their own servers, or even charge money for a product powered by it. OpenAI's big models, on the other hand, tend to keep you locked into their proprietary environment, so that sense of freedom attracted a lot of folks to QVQ early on. Of course, while QVQ's team was refining the model, the AI scene exploded with new approaches, particularly around large reasoning models LRMs. This wave included OpenAI's O3 series and DeepSeek R1 from High Flyer Capital Management's DeepSeek Lab in Hong Kong. According to a web analytics firm called SimilarWeb Deepseeks R1 soared to become the second most visited AI model providing website, coming in behind only OpenAI. That's huge. So we ended up in a landscape where scaling up to crazy sizes is one popular way to push performance. But there's also a new interest in making these models reason better by adjusting how they think through each answer. Qalibaba unveiling QVQ32B with a big emphasis on reinforcement learning to scale up a model's reasoning instead of only counting on bigger and bigger parameter counts. So let's get into how this model is set up. It has 32.5 billion parameters, with 31 billion of those being non embedding parameters. It uses a standard causal language model architecture with 64 transformer layers. It also incorporates rope for positional embedding, swig glue, RMS norm, and an attention QKV bias. There's a generalized query attention scheme that uses 40 heads for the query and 8 for key value. The context length is enormous at 131,072 tokens, which is a big step up from many older models that get stuck at 2,000 or maybe 4,000 tokens. Because it's built on Quin 2.5 code, some folks had to update their libraries to at least transformers 4.37.0 or so to avoid errors. We'll get back to this in a moment, but since AI is evolving fast, knowing how to use it can make a huge difference. That's why we've partnered with Growth School to bring you something valuable. 2024 and early 2025 have been unpredictable. Jobs come and go and financial security is never guaranteed. That's why multiple income streams aren't just smart, they're essential. AI can help, and with the right skills, you could earn an extra $10,000 a month. Now, if you're wondering how to get started growth something really cool. They're offering a three hour hands on AI training where you'll learn to use over 25 powerful AI tools. Normally it's paid, but the first 1000 AI Revolution viewers can join for free using the link in the description. On top of that, you'll get $500 worth of bonus resources just for signing up. The training covers everything job hunting tips, salary negotiation, mastering Excel, even content creation. And it's not just for tech experts. Whether you're in finance, sales, marketing, HR, or even still studying, this can work for you. Growth School has already helped millions of people level up and this could be your turn to stay ahead in an AI driven world. So if this sounds like your kind of thing. Hit the link below to grab your free spot. Plus, don't miss joining Growth School's WhatsApp community. It's a great place to connect with others. Diving into AI too All right, now Back to QVQ of 32B and why it's got everyone talking. The Qin team recommends so certain sampling parameters if you want the best results, such as a temperature of 0.6, top P of around 0.95 and top K somewhere between 20 and 40. They also say you can enable something called yarn scaling if you're going to toss more than 32,768 tokens at the model because that helps it handle huge amounts of text more effectively. But be warned that if you're using vllm, you'll only get static yarn right now, so the scaling factor remains fixed no matter how long your input is. The real secret sauce is how the QEN team used reinforcement learning in two distinct phases. In the first phase they focused on math and coding. They had a system that checked whether a final solution to a math problem was correct and used a code execution server to see if code would pass certain test cases. Only if the solutions or code actually worked did they reinforce the model. After that, they gave it a second phase of training using general reward models and rule based verifiers, and that step was designed to boost performance on everyday tasks, alignment with human preferences and agentic capabilities so it wouldn't just be a one trick pony stuck doing math or code. According to them, these extra RL steps didn't wreck any of the math and coding gains. Instead, they claim that QVQ32B is an all around better problem solver with improved reasoning across the board. Now the talk of the town is how this AI model stacks up in benchmarks compared to Deepseek R1 plus a few others like O1 mini. According to the stats shared in one of the more official references, QVQ32B clocked 79.5 on AIME24, where DeepSeek R1 apparently gets 79.8 63.4 on Live Code bench close to DeepSeek R1 65.9 73.1 on Live bench versus 71.6 83.9 on IF Eval compared to 83.3 and 66.4 on BFCL compared to 62.8. BFCL stands for Berkeley Function Calling Leaderboard, which tests a model's ability to call tools or APIs in a structured way. Those are some pretty serious results for a fraction of the parameter size. People are understandably skeptical when a model that is so much smaller can score in that same ballpark. Some suspect maybe there's a bit of selection bias, or maybe QVQ32B is just super optimized for certain benchmarks. But there's no denying that being able to run more comfortably on your typical hardware, or at least on fewer GPUs, while hitting those kinds of scores is a big achievement. Another plus is that it's open source. This means you can download the weights from Hugging Face or Model Scope. The code and license are Apache 2.0, and if you're an enterprise user or a curious hobbyist, you can fine tune it for your own domain or you can keep it offline for privacy reasons. That offline usage angle is a big deal for some companies that have hesitations about sending data to a Chinese e commerce giant's system or any big corporate provider's server. The argument from Qenteam is that you can skip the official QAT interface if you like, pull down the model's weights and run it on your own secure machines with no phoning home. On the community side, folks on Reddit got really fired up about it. Some are calling it a game changer because it's a 32B model that can sometimes outperform a huge moated system, though others caution that real life usage can differ from nice looking benchmarks. A number of testers report that it often thinks at length, generating a huge chain of reasoning tokens that can slow down the time to final answer. That's probably a side effect of the way it's taught to reflect on everything so carefully. Some see that as a feature because it leads to fewer mistakes, while others find it too slow or borderline verbose for practical usage. People are saying you can tweak the system prompts or sampling parameters to rein in that tendency and get more direct answers. Meanwhile, a few have found that if you quantize it down to 4 bit or 6 bit for local inference, it can start messing up certain tasks, though others claim it's still pretty effective for a wide variety of queries. Because the model is new, a lot of folks are still feeling out how stable the performance remains once you compress it to smaller footprints. From an enterprise perspective, you can see the appeal. CEOs and CTOs love hearing that they can get near top of the line reasoning capabilities with fewer GPUs or possibly a single machine with a big GPU rather than shelling out for a bigger cluster. The Quen team specifically points out use cases in automated data analysis, strategic business planning, financial modeling, software development, and even customer Service. They say QVQ32B can provide more contextually aware responses because it can see up to 131,000 tokens in its input. That means you can feed it entire code bases or large documents and let it figure out relevant connections. Of course, not everyone is going to want to rely on a brand new model from Alibaba, but the fact that it's open weight means you can host it on your own infrastructure, do a ton of domain specific fine tuning and not worry about license fees or usage restrictions. You also get the advantage of seeing the entire chain of thought if you configure it that way so you can track how it arrives at certain answers. Alibaba even released code snippets showing how to run it through hugging face transformers or through Alibaba Cloud's Dash scope API. If you're used to OpenAI style, you'll recognize a similar chat completion method. A typical snippet loads the QVQ32B model, sets up a few arguments for temperature and maxnewtokens, then sends a query. The difference is that there's an optional reasoning output that prints a hidden chain of thought in real time. Many watchers are calling this AI model a milestone in the quest for artificial general intelligence. The Quentin themselves hint that future versions may push even more advanced forms of scaled RL plus deeper agentic capabilities, meaning the model can adapt its reasoning if it observes that something in the environment, for example a search result or a user's feedback contradicts an earlier assumption. They're also looking at ways to make RL more efficient so it can be scaled even further. The dream is that eventually you won't need half a trillion parameters if the model can do enough reasoning to compensate for a smaller knowledge store. There's still a big question mark over how well that logic stands up when the model is asked for obscure facts or domain knowledge. Though something as large as deep seq R1 might contain more raw data about specialized topics. That's where retrieval augmented techniques can fill in the gaps for a smaller model. The Qin team absolutely acknowledges that no single approach is a silver bullet for them. QVQ32B is a proof that reinforcement learning done in multiple stages can coax out a deeper intelligence from a more modestly sized base model. Another detail that's been repeated is the extended context capability. At 131072 tokens, QVQ32B can handle extremely long text. Some users tested it with entire code repositories or massive transcripts, and the model apparently can keep track of a lot more references than older or smaller models. The team calls out that you do have to be mindful when enabling yarn scaling, because the framework you choose might only support a static version of it, which can degrade performance on short texts while letting you handle the bigger ones. But for folks who have a real need to wrangle large documents, it's better than hitting the dreaded context limit. One thing the Quinn team warns is to keep an eye on your gpu, memory usage and thoroughly throughput on Hugging Face. They note that the model is definitely not a small model in absolute terms, so if you only have a single 12 gigabyte consumer GPU, you're going to need to rely on quantization or partial CPU offloading to run it. In any case, the response on social media has been intense. People on X have praised it for being blazingly fast, beating Deepseek R1 and OpenAI's O1 mini, or sometimes outscoring the big gun. Some folks from Hugging Face mentioned how easy it is to deploy on their endpoint with basically one click. Others, particularly in the local LLAMA or local model community, are excited to see if QVQ32B can be an actual competitor to Deepseek for those who want to run everything offline. Not everyone is convinced, though. Some folks note that it might do well on certain math tasks, but not on specialized integrals or advanced logic puzzles. Another chunk of the community complains that it basically overthinks and dumps tens of thousands of tokens in a single response, which can be cool for thoroughness but tough if you're in a hurry. There's also been repeated mention that real usage might not precisely match the official benchmarks, which is something we often see in AI model announcements. Still, if you step back, you have to give credit to the QEN team for pushing the conversation forward and making a strong case that multi stage RL training can do wonders. The fact that you can get this model free under the Apache 2.0 license is a big plus for developers who want to test, Tinker, or commercialize. At the very least, it represents a step toward bridging the gap between massive parameter counts and efficient reasoning. Whether it truly stands toe to toe with Deepseek R1 in every scenario is something that only broader testing will confirm. But the hype is real, and for good reason. It's a model that can definitely handle a wide range of tasks from writing code to solving math to generating coherent multi paragraph explanations, all while letting you keep it on your own hardware. If you're an AI enthusiast, it's worth giving it a spin either by testing it on hugging face on ModelScope or by checking out Quenchat if you'd rather not mess with the code. It's exciting to see these open models come out with advanced reasoning capabilities and not require specialized super computing resources. Whether you agree with the self reported benchmarks or not, QVQ32B is out there right now so you can test it with your own prompts, measure how it compares to your existing solutions, and decide for yourself if it holds up. If it really can stand toe to toe with Deepseek R1 in many tasks, we might be looking at the next big shift in the AI world where reinforcement, learning and smaller parameter counts come together to challenge the huge behemoths. One thing is for sure, we're not going to get bored in this domain anytime soon. Drop your thoughts in the comments and if you enjoyed this breakdown, don't forget to like subscribe. Subscribe and share. Thanks for watching and I'll catch you in the next one.