Transcription
Hello everyone. Today we have something fresh from the world of AI for programmers to analyze. We came across some materials, more precisely a video review from the World of AI channel about the new Deepsig V3.1 Coder model. And the task is to figure out what kind of beast this is, uh, what makes it interesting. Especially, of course, it's interesting to compare it with competitors, well, say, with Cloud ST 4, and to understand how useful and accessible all this is for us developers. Let's dig in. >> Yes, hello. The model and PR made a bit of a splash, they claim all sorts of interesting things, and the speed seems to be higher, and some agent capabilities have been upgraded, and this hybrid reasoning mode. Well, let's see what's actually in the review, whether these claims are confirmed. >> Uh-huh. So, deepsк V3.1. has officially come out. And right away, this hybrid mode sounds like they're trying to, you know, respond quickly and think when needed. Something similar seemed to have already existed, I think, with ContriQ. Or is there something new here. >> Correct, it seems so. But here, as I understand it, it's precisely a refined idea, the hybrid mode, yes, that's one of the main features. The point is for the model itself, so to speak, to choose a strategy. If the task is simple, the answer is quick, without much thought. But if it's something tricky, the reasoning mode is activated to make the answer higher quality. The source actually compares it with Quentin 3, but they say that here, firstly, tool calling has been significantly improved. This is when the model needs to communicate with the external world via APIs, and secondly, the model itself has allegedly become, well, much faster than its predecessor Deepsek R1, especially in this reasoning mode. So, yes, they are trying to combine speed and depth. Understood. So, what else? Context window 128,000 tokens. In a world where everyone is measuring themselves by these tokens, 128,000, what is that like? Normal, little? >> Well, 128,000 is, I would say, quite solid. For most programming tasks, this will be enough. You can throw in a large piece of code, analyze extensive documentation, or have a long dialogue during development without losing context. This is a quite practical size. >> Uh-huh, >> of course, if you compare it with giants like Geni, which claim a million or even more, then 128K is not outer space. But here there's always a compromise, you understand? Huge windows are often slower and more expensive. And Deep ПСК seems to be aiming for efficiency for coding, for agent scenarios. So 128K is most likely a conscious choice. The golden mean, perhaps. Logical, okay, that's clear. But the main hype, it seems, is around performance specifically in code. The source says directly: "It can compete with PLM Sonet 4. And PLM Set 4 is one of the leaders for coding right now. >> And at the same time, DeepsК is cheaper and possibly can even be run locally. Now that's really interesting. >> Precisely, positioning as an alternative to SIT 4 is very ambitious, and they back it up with test numbers. The review mentioned, I'll say now, uh-huh, 66 points on their bench verified. This is a benchmark where they check how correctly the model solves practical coding tasks, you know, on real examples. And another 313 points on Terminal bench. This is about working with the command line. >> Well, numbers are numbers, of course. But what's behind them? Some qualitative improvements. >> Yes, and they talk about that too. It is emphasized that multi-step reasoning has been strengthened. This is precisely what is important for complex things. Well, debugging tangled code, refactoring a large project, finding non-obvious errors. They also talk about improvements in the thinking mode, that is, when the model digs deep. And what is really important for modern applications is more efficient use of tools and execution of multi-stage agent tasks. Agent tasks, you mean when AI doesn't just write code, but performs a sequence of actions, right, like finding information itself, using APIs itself. >> Exactly, not just a code generator, but rather an assistant that can take on a whole piece of work, interaction with other systems. It seems the model was designed for this. >> So, what about the money. If you believe the review, how much cheaper is it compared to Son 4? Noticeably. The numbers they gave were: for input tokens, that is, for what you feed the model, 7 cents per million tokens, if the data is already in the cache. This is called a cache hit. >> And 56 cents per million if the data is new. Cache miss, and for output tokens, that is, for the generated response, a dollar 68 per million. >> Hmm. A dollar 68. And how much is Sony 4 for comparison? Well, the exact prices for Sony 4 can vary, but by general estimates, PSIC V3.1 turns out to be significantly cheaper, noticeably cheaper, especially for large volumes. This is, of course, a very big plus for startups, for teams that run AI a lot and for whom API cost is a critical factor. >> Yes, price decides a lot. And here's another point that was noted in the review - support for Entropic API. A very smart move, I think. >> What does that mean? They just made their API similar to the API from Anthropic, the creators of Claude. >> Exactly, they added compatibility. That is, if you already have code that works with Claude through its API, you can, well, almost without changes, just plug in Deepsik and try it. They are essentially telling developers: "Guys, try our model. It can do roughly the same as Son 4, and in some ways, especially in agent tasks, it might even be better. And it's cheaper, and faster, and you don't even have to rewrite almost anything." >> Clever. That's a direct way to enter the market and poach users. >> Absolutely, it's a bet on price, speed, and this optimization for agent tasks. Plus, easy entry for those already in the Anthropic ecosystem. It all sounds very tempting, but for now it's theory and marketing. But how is it in practice? The review had examples. What did they show? I remember something about Doom. >> Yes, there was an example with a Doom clone. The story is like this. They took code that was originally generated by Cl, but it was apparently not ideal. And Deeps V31 refined it. And it is claimed that it handled tool usage and these multi-step tasks perfectly, without errors. >> So it managed to correctly integrate different parts, call the necessary functions, and assemble a working game prototype. A clone without errors from the first try. Well, that sounds strong. >> Well, from the first try is debatable. It was code refinement from Claude. But the very fact that the model could correctly handle a complex task with tool integration confirms the claimed improvements in reliability and agent capabilities. The result in the video looked quite playable, by the way. >> Interesting, what else was there? Seems like some butterfly. Uh-huh, there was generation of an SVG butterfly. This is a vector graphics format. They tested it through different tools. Killow Code, Any Coder, and just in a chatbot. >> And everywhere a cute colorful butterfly turned out. And what's funny, sometimes even animated, although they didn't specifically ask for animation. >> Oh, animated. It thought of it itself, >> apparently, yes. Or the platform settings influenced it. But in AnyCoder, the result was even more interesting. A whole mini-application with buttons was generated. You could control the animation and color of this butterfly. >> Hmm. So the model didn't just draw a picture, but also made an interface for it. >> It turns out so. This shows flexibility, that it can not only write code but also generate visuals, following certain criteria like symmetry, and even add interactivity. The fact that different platforms gave slightly different but interesting results speaks to its potential. >> Next on the list of examples. Something about crypto. Frontend for an exchange. >> Yes, there was an example of creating a frontend for a decentralized crypto exchange. Dex through Any Coder, and they particularly noted the speed. They said that about 800 lines of code were generated very quickly. >> 800 lines is a lot, and speed is really that important. Well, you can wait. >> It's very important in prototyping, when you're just exploring an idea, want to quickly assemble something, show it, test it. Here, the ability to get 800 lines of a working framework in minutes, not hours, really changes things. It allows for faster iteration, hypothesis testing. Judging by the review, the result looked like a standard interface for such an exchange. That is, as a starting point. Excellent. >> Okay, understood. And there was, I think, the most complex example. Some financial application. >> Yes, this is perhaps the most impressive example from the review. A financial application for tracking cryptocurrency and stock prices in real-time. It was made through Killow Code. What's interesting, they said the cost of generation was about $1. Well, true, using the free credits that Kill Code gives upon registration. >> A dollar for a whole application. Not bad. And what could it do? >> And here's the most interesting part. The result, as described, was not just a set of pages, it was an application with working APIs for obtaining real-time market data, i.e., with real quotes. There was clean logic in the backend and a frontend that displayed this data and updated it. >> Wow. So the model itself found the necessary APIs, connected them, wrote the backend, wrote the frontend that communicates with this backend, >> judging by the description. Yes, this is precisely the demonstration of those agent capabilities and working with tools. The model didn't just write code separately, it understood a complex task that required external data, found a way to get it itself through APIs, structured it all into a working application. The reviewer called the result exceptional and noted that the data integration looks very natural. >> And all this in one pass, as they say. Well, at least that's what the review claims, that they gave the task and got a ready-to-use working application. If that's really the case, then this is a very serious level. Seamless work with live data via API is the key here. >> Yes, it sounds powerful. And that even games, maybe something about Minecraft. >> And even a Minecraft clone was mentioned. Of course, not a full-fledged game, but a basic prototype, but with the ability to place and destroy blocks. They said it looked quite decent for something generated automatically. >> Well, that's interesting. So, the ambition extends to Game Dev, >> well, at least to prototyping in game development. Minecraft is seemingly a simple thing, it has a 3D space, object interaction, some physics, game logic. Handling such a thing is a good stress test for the model. It shows that it can grasp spatial concepts and interaction rules. Even if the result requires refinement, the very ability to create a working foundation for such a project is serious. The examples are certainly impressive, from Doom to Fintech and Minecraft. >> A logical question: how can an ordinary developer get their hands on this model? Where can they get it? How to try it? >> The review listed the options, >> yes? And there seem to be quite a few options. That's also a plus. So, what were they, >> yes, and there seem to be quite a few options. That's also a plus. So, what were they? First, the official Deepsi chatbot. Well, that's the easiest way. Just go to their website and chat. >> Uh-huh. Like with chat GPT or CLD. >> Something like that, yes. Second option is to use the API. We've already talked about Anthropic API support. That is, if you're already working with Claude, you can try Deepsik through the same interface. Third. Dips also has its own native API. It can be used directly or through aggregator platforms like Open Router. Open Router is where you can call different models through one API. Yes, exactly. It's convenient for comparison or if you want to switch flexibly. And fourth, the most interesting for us, is integration with coding tools. >> Oh, that's important. What IDEs or plugins has it been integrated into? So, the review mentioned them. It seems you can choose deeps v3, but the latest 3.1 might be through local launch using the together AI API, meaning not directly out of the box yet, maybe. >> Hmm, locally via API. Well, that's something. >> Next, Cloud Code. This, I understand, is some kind of cloud environment. You need to configure environment variables a bit. And then you can use DeepS 3.1 as an alternative to SIT, and, as they say, a faster and cheaper one. >> So, also >> they used it for the butterfly and crypto frontend examples. They noted that it's very fast, but there might be limits on free usage. Then you might need tokens from Haging Face. >> Understood. A premium model >> looks like it. Next, Kill Code. The review directly recommended it because upon registration they give free credits, and you can test the Deepsek 3.1 API with them. They used it for the butterfly and that financial application examples. >> Uh-huh. So Kill Code is a good start to try the API for free. >> It turns out so. And Klein was also mentioned. It was called one of the best ways to use it, but without much detail about what it is and how it works. Yes, there's a choice. From a simple chatbot to integration into various tools. And, importantly, you can try it for free or almost free in many places. >> Exactly. Accessibility is clearly one of the strong points, uh, that they are betting on. It's easy to start experimenting. >> Okay, let's try to put it all together then. Based on this review. What's the conclusion? Deepsek 3.1 is a serious player in the market and for coding. A real competitor to the top ones. The main advantages, as I understood, are lower price, higher speed, better focus on agent tasks. So >> overall, yes, but the reviewer gives a, you know, balanced assessment. He doesn't say that Deepsк beats everyone in code quality. He uses the phrasing that the code might not always surpass Sonet 4. That is, there might be cases where Sonet 4 generates something more elegant or optimal. >> Uh-huh. So, not an ultimate killer of Sonet 4 in terms of code quality. >> Ultimate. But then the author adds that the code from Deeps 3.1 is definitely decent, reputable, and the model is capable of competing with Sona 4. So it doesn't mean the code is bad. It's good, functional, and in many cases will be no worse. And if you consider other factors, >> these other factors, price, speed, these agent things, tool integration, multi-step capabilities. >> Exactly, the review focuses on the combination of these factors. Very attractive price, high task execution speed, advanced capabilities for creating AI agents, and, importantly, potential availability for local launch, although this needs to be verified. This combination makes the model a very interesting alternative. Not necessarily superior in everything, but a strong alternative. And again, it's easy to try it yourself and decide if it suits your tasks. >> Well, the picture is becoming quite clear. Judging by this review, V31 is indeed a powerful and, importantly, accessible tool. It's definitely worth paying attention to for those who write code and are interested. And this combination of speed, price, focus on agent capabilities, plus being backed by examples like a financial application with real data, looks convincing. >> Exactly. The examples show that it's not just a toy, but a tool with serious practical potential. From game prototypes to complex applications with integration of external services. Of course, as with any AI, the devil is in the details. Real effectiveness will depend on the specific task, on how the prompt is formulated, how the interaction is configured. You need to try it on your own projects. >> Absolutely. >> But the main thing is, such an opportunity exists. The model is available, it can be tested, compared, and developers can decide for themselves how well it suits them. And, you know, while we were discussing this, an idea came to mind: if models like Deepsek 3.1, powerful, fast, cheap, focused on writing and integrating code, are becoming more accessible, how will this change development itself? How close are we to the moment when you can just describe your idea in great detail, and the AI agent will write the code, debug it, connect all the components, and deliver a finished product? >> Good question. It seems we are moving in this direction, not tomorrow, of course, but models like DeepsК 3.1 are a clear step towards more automated development, where the human role shifts from writing every line of code to setting tasks, control, and architecture. >> Interesting times are coming. Well, with that, perhaps our today's analysis of materials on Deeps 3.1 coder can be considered complete. Thank you to everyone who listened. >> Yes, it was interesting to dig into. Thank you.