Transcription
Deep just quietly, without loud announcements, rolled out the V3 and1 model. And while some talk about a breakthrough, others see only a minor update. But the truth is, almost everyone is missing the main point. Behind this release lies something more than just new numbers in benchmarks – it's a cold geopolitical calculation. Today, we will break down this release into atoms. I will show why the debates about base and Instruct models were just a smokescreen. What is the real power of V3 and1 and how is it connected to the new Chinese chips from Huawei? By the end of this video, you will understand whether V3 and1 is a true competitor to CL 4 and Gemini 2.5, and why its greatest strength is also its biggest limitation. And we will start with the confusion that Deep created intentionally. The whole story began with the quiet appearance on Hugging Face of a model named Deepsek V3 and1 Base. The community immediately divided, what is this, why Base? A base model is, in essence, raw material. It's a neural network after the main training stage, but before it's specifically trained for dialogue and instruction following. It's not intended for communication. Trying to have a dialogue with it, especially on local hardware, is an almost guaranteed disappointment. And that's where the storm began. Why release a semi-finished product? Is this all the update? But this was only the first part of the plan, because the real release happened literally a few days later, on August 21st. And it was a completely different model, a full-fledged instruct version, ready to go out of the box. And this is where it gets interesting. This is not just an update, it's a solution to the main pain that plagued users of DeepS's previous top model. The key innovation of V3 and1 is its hybrid architecture. Within one model, there are now two modes: normal chat and a deep reasoning or thinking mode. Previously, reasoning was handled by a separate, legendary, but unbearably slow model, R1. It could produce brilliant answers, but it did so so slowly that you could have time to brew coffee and read a chapter of a book. It thought about everything forever. V3 and1 solves this problem. The reasoning mode in the new model works significantly faster and even surpasses R1 in coding and logic tasks. Users confirm that the speed in Think Mode is noticeably higher, and token consumption to achieve results is lower. How does it work? Switching between modes is done using a special token. This means that developers can dynamically control the depth of the model's thought directly within a single request. The company itself calls this the first step into the era of agents. And these are not empty words. The model was specifically created for agentic use and calling external tools. It is intended to be the foundation for creating complex automated systems that can not just answer questions, but perform multi-step tasks. And to prove the seriousness of their intentions, PSI has invested colossal resources into this model. Let's look under the hood, what's behind this update. Firstly, massive fine-tuning. On top of the already existing V3, the model was further trained on a staggering volume of 840 billion tokens. The main emphasis was placed on long dialogues and complex instructions. To improve the quality and stability of responses, there were two phases: context expansion: up to 32,000 tokens on 630 billion tokens, and then expansion up to 128,000 on 209 billion tokens. And the result was not long in coming. In benchmarks, V3 and1 shows a colossal leap. On one of the tests, performance increased by 20 points, reaching the mark of 66. In a complex test on verified code SVBCH, the model scored 66%, while its predecessor R1 scored only 44.6%. But the main triumph is in the Aider benchmark, where V3 and1 showed a result of 71.6%, surpassing even Cloud Opus 4. This makes it one of the best, if not the best, open models for coding today. In reasoning tasks, the model also improved, scoring 60 points in the Artificial Analysis index, surpassing R1 with its 59 points. However, here it still conceded to some competitors. But the intriguing detail is hidden not in the benchmarks, but in the technical specifications. The model was trained using the new U and8 MOI FP8 format. And here we move from technology to geopolitics, let's call it the Huawei Gambit. In one of the announcements, there was a phrase that this format was developed specifically for the new generation of Chinese-made chips that will soon enter the market. This refers, in particular, to Huawei Ascent 920S chips. Some are even speculating that it is preparing the ground for the release of its own specialized ASIC chip. Think about it. This is no longer just the creation of a new neural network, it's the construction of an entire vertically integrated ecosystem of software and hardware, independent of Western technologies. The quiet release of V3 and1 turns out to be a strategic move in the global chip war. Now that we know the power and strategy hidden within this model, let's come down to earth and see how it looks in the real world compared to market titans. Let's start with the context window. Deep has increased it to 128,000 tokens. This is good, but not a record. Gemini 2.5 Pro and Clot 4 have long been operating with contexts up to 1 million tokens. Here, Deep PSI is clearly not trying to chase numbers, but is taking a pragmatic position: 128,000 tokens are enough for most engineering tasks, and this allows for a balance between capabilities and cost. But where V3 and1 still lags is in advanced agent functions. And this is perhaps its main drawback today. Despite all the statements about the era of agents, the model does not support Function calling directly in reasoning mode. This is critically important. This means that the model cannot simultaneously think deeply and interact with external tools, get data via API, work with databases, run scripts. It can either think or act. Competitors like CLD 4 with its Extended Thinking mode or GPT are a step ahead here, as their architecture allows combining these processes. This seriously limits V3 and1 in building truly complex and autonomous AI agents. Now let's listen to what real users are saying. The picture is very mixed. On the one hand, the model is praised for what it can actually do. The code quality, according to reviews, is at the highest level. Some developers note that the code turns out smoother than GPT5's, and the model often solves complex tasks on the first attempt. The speed and efficiency of the reasoning mode compared to R1 is something that almost everyone who suffered from its slowness before thanks the model for. The transparency of thought, when the entire chain of reasoning is visible, is also a huge plus for debugging and understanding AI logic. But on the other hand, there is also serious criticism. Some users complain that the model has acquired an unpleasant trait. It has become sycophantic, trying to please, even if it goes against facts. The number of hallucinations has increased, and sometimes the model switches to Chinese for no reason. There are also complaints that the model has become worse at following formatting and text volume instructions. Its response style, according to some, has become suspiciously similar to G54, which fuels rumors of data distillation. So, who is this model for? First and foremost, for developers, engineers, and analysts. For those who have to work daily with huge code repositories, analyze long logs, and decipher multi-page documentation. If you want to run it locally, be prepared for serious requirements. By estimates, you will need at least 128 GB of RAM and several powerful video cards. And for truly comfortable work, 256 GB of RAM. So, what's the final verdict? Deeps V3 and1 is not a killer, Claude, or Gemini. It's a powerful workhorse and a strategic asset. Its strength lies in balance, speed, and specialization for specific engineering tasks and future hardware ecosystems. It's a surgeon's scalpel, not a Swiss Army knife. It shows top results in coding, surpassing even some closed giants. But its limitations in agent scenarios do not allow it to be called a universal solution. It's an evolutionary step that prepares the ground for the future, possibly for R2, which is expected in the fall. And now you look at all this: Gambit, Huawei, my architecture, gigabytes of RAM, and ask yourself one question: "What do I do with all this? How can I apply it to my business tomorrow, not in a year?" And this is where it gets interesting. Instead of trying to tame another giant, you can use ready-made, honed mechanisms that solve your business tasks here and now without unnecessary complexity. We call this micro-automations. These are not clumsy prompts or universal models. These are small, ready-made mechanisms that quickly solve clear, understandable tasks in your business. Imagine that you no longer write prompts. Instead, you answer a few simple questions: what is the product? Who is your audience? What are their doubts? Then you press a button and immediately get a ready, tested result that doesn't need manual refinement. Why does this work so reliably? Because inside each micro-automation, the experience of the best market experts is already packaged. We take specific methodologies and approaches that have brought real money and leads hundreds of times, and carefully package them into a mechanism that you can use again and again. What exactly can be solved today with micro-automation in Neurobuilder? For example, you can create an offer in minutes that will immediately attract clients. You can conduct a deep audience analysis that used to take days, or create a ready-made email sequence that will lead people to purchase tomorrow. Here are just a few tasks that are easily solved in the club. Creating an offer after which people immediately register for a webinar or consultation. Developing an email sequence that turns cold subscribers into clients in a week. Generating a content plan for a month ahead that brings leads every day. Designing a landing page that immediately starts generating profit. Writing scripts for short videos that grab the audience from the first seconds. Conducting a quick and accurate competitor analysis to immediately see their strengths and weaknesses. Creating viral content that people themselves repost and discuss. This is no longer abstract theory, these are concrete and proven solutions. And here you face a simple choice. You can continue searching for the perfect prompt and spend time on manual work. Or you can use ready-made mechanisms right now that have already passed hundreds of tests and give guaranteed results. If you choose the latter, then the closed club Neurobuilder is created precisely for you. Dozens of micro-automations that solve your tasks quickly and efficiently are already gathered here. Think about how your work will change when you no longer have to reinvent the wheel every time. If you are interested in this approach, follow the link in the description and join. We offer you not another course, but a ready-made, working solution for your business. Yeah.