Model

Google just changed the rules of the lightweight model

by

Now, in the race to lead the development of artificial intelligence, something unusual has just happened. Gemini 3 FlashGoogle’s new model, has surpassed GPT-5.2 Extra High, the higher-reasoning variant of OpenAI, in several performance tests. And that forces us to rethink some of the rules that we took for granted. A fast model that also reasons. Google’s new model comes with a very specific promise: to demonstrate that “speed and scalability do not have to come at the expense of intelligence.” Although it has been designed with efficiency in mind, both in cost and speed, Google insists that Gemini 3 Flash also excels at reasoning tasks. According to the company, the model can adjust your thinking ability. It is able to “think” for longer when the use case requires it, but it also uses 30% fewer tokens on average than Gemini 2.5 Promeasured with typical traffic, to complete a wide variety of tasks with high precision and without penalizing response times. The truth is in the benchmarks. Are the benchmarks perfect? No. But they are still one of the most useful tools we have for comparing AI models.confront them against each other and detect in which scenarios they perform better or worse. And in this area, Gemini 3 Flash comes out well. In SimpleQA Verifieda test that measures reliability in knowledge questions, Gemini 3 Flash achieves 68.7% compared to 38.0% for GPT-5.2 Extra High. In multimodal reasoning, within MMMU-Pro, Google’s model scores 81.2% compared to OpenAI’s 79.5%. In Video-MMMU, Flash achieves 86.9% compared to 85.9% for GPT-5.2 Extra High. If we look at multilingual and cultural capabilities, Flash is again ahead, with 91.8% compared to 89.6% for GPT-5.2 Extra High. In Global PIQA, focused on common sense in 100 languages, the difference remains: 92.8% for Flash versus 91.2% for the OpenAI model. Everything indicates that Gemini 3 Flash is specially optimized to capture nuances outside of English and reason more fluently in global contexts. He also excels in the use of tools and agents. In Toolathlon, Flash scores 49.4% compared to GPT-5.2 Extra High’s 46.3%. In the FACTS Benchmark Suite, the difference is tighter, but still in favor of Google: 61.9% versus 61.4%. In long-term tool execution tasks, Flash appears to show greater consistency. But he is not the king of pure reasoning. Now, it is worth looking at the complete photo. Although Gemini 3 Flash outperforms the best OpenAI model in several tests, if you are looking for “pure” reasoning, the balance changes. In the most demanding tests in this area, GPT-5.2 Extra High continues to set the benchmark. OpenAI’s model leads ARC-AGI-2, focused on visual puzzles, with 52.9% compared to Flash’s 33.6%. In AIME 2025, with code execution, it reaches 100% compared to 99.7%. And in SWE-bench Verified, aimed at software engineering, it obtains 80.0% compared to 78.0% for Gemini 3 Flash. What exactly is GPT-5.2 Extra High. Throughout the article the name GPT-5.2 Extra High appears several times, and it is normal to wonder if it is something new or little known. In reality, it is not a model that is usually mentioned to the general public. Google uses this designation in its comparison table to refer to the maximum level of reasoning available in the OpenAI API for GPT-5.2 Thinking and Pro. In the official OpenAI documentation it is identified as “xhigh”. Where you can use Gemini 3 Flash. Access to Gemini 3 Flash is not country dependent. If you have access to the Gemini appyou are already using this model, which has become the default option. It is also reaching developers through the API, AI Studio and Vertex AI. In the United States, the deployment goes a step further, as the Gemini 3 Flash has become the default model of the AI Mode of the Google search engine. The price of using Gemini 3 Flash. For those who want to integrate Gemini 3 Flash into their applications, the model costs $0.50 per million input tokens and $3 per million output tokens. This is a slight increase over Gemini Flash 2.5, which was $0.30 per million tokens in and $2.50 per million tokens out. An increasingly tight race. Gone are the days when Google tried to confront ChatGPT with Bard, or when OpenAI seemed to be years ahead of the rest. Today, the distances between the big players in AI have been drastically reduced. The competition is more direct, more technical and, above all, much closer. Images | Google In Xataka | Amazon is preparing an investment of 10 billion in OpenAI because if you can’t beat your enemy, the best thing is to join him

DeepSeek has launched its new reasoner model. It’s free and beats GPT-5

by

DeepSeek has introduced DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. They are AI models that combine complex reasoning with the ability to use tools autonomously. Why is it important. The company of Hangzhou claims that DeepSeek-V3.2 matches the performance of GPT-5 in multiple reasoning tests. The Speciale model It reaches the level of Gemini-3 Pro and has achieved gold medals in international mathematics and computer science Olympiads. The context. DeepSeek surprised the world in January with a revolutionary model for efficiency and cost. Now it ups the ante with open source systems that throw down the gauntlet directly to OpenAI and Google in reasoning capabilities. Technical innovation. DeepSeek-V3.2 integrates “thinking” directly into tool usage for the first time. You can reason internally while running web searches, operating a calculator, or writing code. The system works in two modes: With visible reasoning (similar to the reasoning seen in ChatGPT and company). Or without any reasoning. The chain of thought persists between tool calls and is restarted only when the user sends a new message. How they have achieved it. Researchers have developed ‘DeepSeek Sparse Attention (DSA)’, an architecture that greatly reduces the computational cost of processing long contexts. The model maintains 671 billion total parameters but activates only 37 billion per token. In figures. DSA cuts the cost of inference in long contexts by approximately 50% compared to the previous dense architecture. The system processes 128,000 context windows tokens in production. Reinforcement training has consumed more than 10% of the total pretraining count. The team has generated more than 1,800 synthetic environments and 85,000 tasks to train agent capabilities. The results. DeepSeek-V3.2-Speciale has won a gold medal at the International Mathematical Olympiad 2025, the International Informatics Olympiad 2025, the ICPC World Finals 2025 and the Chinese Mathematical Olympiad 2025. Both models are available now. V3.2 works on app, web and API. V3.2-Speciale only by API, at least for now. Between the lines. DeepSeek has published the full weights and technical report of the training process. This transparency contrasts with what large American technology companies usually do. Even those that offer open source models such as Call, with an asterisk. The Chinese startup wants to demonstrate that open source systems can compete with the most advanced proprietary models. And it does so while continuing to reduce costs. Yes, but. The benchmarks Public settings do not always reflect performance on real-world tasks. Direct comparisons with GPT-5 either Gemini-3 Pro They depend on specific metrics that may not capture all relevant dimensions. Furthermore, the integration of tools in reasoner mode still needs to be tested in complex real-world use cases. The reduced cost is not as important if the quality of the responses does not hold up. In Xataka | DeepSeek Guide: 36 Features and Things You Can Do for Free with This AI Featured image | Solen Feyissa

It is surely the best model for programming, but it still has a big problem

by

Anthropic has announced Claude Opus 4.5, its most advanced AI model to date. The company claims it is the best in the world for programming, intelligent agents and computing usage, beating OpenAI’s GPT-5.1 Codex-Max and Google’s Gemini 3 Pro. It has also arrived a few days after both as well as Grok 4.1. The general overview. The new model has achieved 80.9% accuracy in SWE-Bench Verified, the benchmark reference to evaluate software engineering capabilities. Anthropic has also put it through its own hiring test for engineers – notably difficult, with a two-hour limit – and the model has outperformed every human candidate who took it. Why is it important. This release solidifies Anthropic as a leader in AI tools for programming. Even Meta uses Claude for its internal Devmate code assistantdespite competing directly with the company in other areas. The improvements are not limited to the code. Opus 4.5 stands out in: Creation of documents, spreadsheets and professional presentations. Deep research tasks with multiple sources. Advanced visual and mathematical reasoning. Management of subagent teams for complex multi-agent systems. In figures. Additionally, Anthropic has drastically reduced the price of its API: from $15/75 per million tokens entrance/exit at 5/25 dollars. And the model is more efficient than its predecessors: In medium effort mode, it equals the performance of Sonnet 4.5 but consumes 76% less tokens. In high mode, it beats Sonnet 4.5 by 4.3 percentage points using 48% less tokens. The context. The company has introduced that “effort” parameter (low, medium, high) that allows developers to control how long and tokens invests the model in solving a problem. It is a trend that OpenAI has also adopted in its latest modelsseeking efficiency without sacrificing quality. In detail. Along with the model, Anthropic has updated its development platform and consumer applications: Claude Code Improves your planning mode: Asks clarifying questions before creating an editable execution plan file. As seen with the Deep Research on duty. Claude for Chrome is now available to all Max users (around $100-$200 per month depending on limits), allowing AI to manage tasks across multiple browser tabs. Claude for Excel opens to Max, Team, and Enterprise users, with support for charts, pivot tables, and file uploads. Endless conversations– Long conversations no longer run into context window limits thanks to automatic summaries. Yes, but. The big problem with Opus 4.5 and Claude in general is its usage limit. Even for Pro and Max subscribers of the first levelthe tokens They sell out quickly. They take five hours to restart from the first message sent. The Opus model, being the most powerful, is also the one that consumes the quotas the fastest. This is the main source of frustration for users who pay $20 or even $100 a month. Anthropic has slightly increased the limits for Max and Team Premium, but the experience is still far from what is expected in a service of this category. Between the lines. The release of Opus 4.5 restores balance to the Anthropic model family. For the past two months, Sonnet 4.5 was outperforming the older Opus 4.1, leaving little reason to use the more expensive model. Now, with three clearly differentiated models (Haiku, Sonnet and Opus), each one has a specific purpose in terms of cost, speed and capacity. And now what. Anthropic follows a clear strategy: position itself as the premium provider for knowledge professionals and developers, competing directly with OpenAI and Google in the field where accuracy and reliability matter most. But if you don’t solve the problem of usage limits, you risk frustrating the very users who could get the most value from the model. In Xataka | AI is transforming the relationship we have with our own ideas: we no longer create, we just “edit” ourselves Featured image | Anthropic

What are the news about Google’s new artificial intelligence model?

by

Let’s tell you what they are the main news of Gemini 3the new version of the model artificial intelligence announced by Google. We already have the first data on its main characteristics. As always, the flashes will go to Gemini 3 Pro, which will be the most advanced version. Here, one thing you want to know is that You will notice few of these new developments when you are using Gemini in a conventional way. Most of these changes are aimed at advanced users. Gemini 3 news A step forward in all areas: Google has presented the results of its model in various types of tests, comparing it with the previous version and its direct competition. He is ahead of everyone in everything, from mathematics to understanding what is happening on the screen or creating code. Reason “at the doctoral level”: That is what the test results also indicate, although where it advances the most is in the mathematical results, with a score of 23.4% for the MathArena Apex test compared to 1.0% for GPT 5.1 or 1.6% for Claude Sonnet 4.5. Integrates with Google Search: Gemini 3 is linked to Google’s AI Mode, integrating into the search engine. Generate visual elements: Gemini 3, has the ability to create interactive visual elements, such as calculators, simulations or widgets in real time. Something especially useful when integrated into the search engine. Sometimes it may not respond to you with text, but with an interactive webapp. More direct answers: Google has fine-tuned the way its model responds, offering more concise responses that offer more valuable information and less flattery, clichés and clichés. Improvements in “Deep Thinking”: Another of the most notable improvements is in deep thinking, in addition to advances in code execution, abstract reasoning and visual understanding. Larger context window: This model has a context window of up to one million tokens, being able to analyze large code repositories or very long texts on which you can later work. Better contextual reasoning: Reasoning is improved, especially with long contexts, to avoid hallucinations. Parallel reasoning improvements: Abilities to reason with visual and textual data at the same time are improved, improving accuracy when interpreting tables, diagrams and interfaces. Improvements in its multimodal mode: The analysis of all types of information is improved. For example, you can decipher and even translate handwritten recipes in different languages, and use them to create a cookbook that you can share. You can also analyze sports matches, scrutinize research data and generate code from it. Programming improvements: As we said at the beginning, one of the biggest improvements of this model is in its ability to program. Improved your agent mode: Your ability to use tools and operate a computer through the terminal using agent mode has also been improved. Agents with Gemini can now autonomously plan and improve more complex software tasks. Gemini 3 will begin to be available in the coming daysalthough as we told you at the beginning, it is possible that many of the differences will not be noticed unless you are going to try to take advantage of them in an advanced way. In Xataka Basics | The best prompts to save hours of work and do your tasks with ChatGPT, Gemini, Copilot or other artificial intelligence

Renault is already pushing for Europe to copy the Chinese model

by

The statements have been as concise as they are clear: “You cannot come to Europe and build four plates with wheels and seats with little added value. What we have to do is commit them to teach us, to come with products with added value. We did not do it like that when we went to China, they should not do it when they come to Europe” The words are from Josep Maria Recasens, president of Renault Spain, and reflect in three sentences the situation that the industry is experiencing in Europe, its internal debates and its fears. Added value. This is what Recasens has demanded at the 1st Automotive Forum, organized by the Automotive Press Group to which it belongs. The Automotive Tribune. The president of Renault Spain, who is also the president of ANFAC (the manufacturers’ association in our country) has demanded that Europe force Chinese brands to associate with European ones so that they “teach us” how they make their products. In Recasens’ opinion, Europe is opening the door to Chinese brands, allowing them to build “four plates with wheels and seats with little added value.” It is a veiled statement that points to the Chinese factories that are settling in our country but that, however, plan to produce vehicles based on kits that already come pre-assembled from China. What do they teach us? When the president of Renault asks that the European Union force Chinese manufacturers “to teach us” it is for two reasons. The first is that China forced foreign manufacturers to partner with their local firms to produce on its soil. What did they earn? Obviously, knowledge. Just take a look at the MG4 Electric to understand the extent to which its partnership with Volkswagen has borne fruit. At the same time, foreign manufacturers could produce at a much lower price and had access to the largest market in the world. What, we assume, they did not imagine is that China was going to surpass the West. Yes, let them teach us. The second point referred to in “let them teach us” is evident: the president of Renault and Anfac recognizes that, at least in part, China is ahead. And the French company itself has gone to Shanghai to develop your Renault Twingoa car whose heart has been created internally in China in record time for the European industry. But there have also been curious situations such as Mazda has brought the Mazda 6e to Europea car developed by Changan in China that, given its success, they have decided to test on European soil with a groundbreaking price per size. And the warnings don’t end there. The industry has entered a fever to shorten deadlines and approaching the times of Chinese development. The consultants warn that, at the level of quality, there is no difference with the Europeans. Others warn Japanese firms that their extreme attention to detail and conservative evolutions they may have left them behind. In question. Recasens’ words also emphasize the misgivings that have arisen among European manufacturers seeing how Chinese companies are arriving on our soil. With the intention of stopping the arrival of Chinese electric cars at knockdown prices, Europe applied variable tariffs to each brand depending on the supposed help they have received from the Chinese Government in the form of soft loans or the transfer of land. The promise is that they would not pay if they manufactured in Europe. But the first factories are also in question. Chery opted for assemble car kits in Barcelona. That is, cars that arrive almost assembled from the other side of the world and to which the final touches are given in the Spanish city. Now, the European Union is studying whether or not the electric Omoda 5 has to pay tariffs by understanding that added value is not being created around the production of said car. But not only Chery. The Chery case was the first but it has not been the only one. Stéphane Séjourné, vice president for Prosperity and Industrial Strategy of the European Commission, has assured the Italian newspaper La Stampa that the institution also has the factory in its sights BYD in Hungary or the plans that CATL has in Europe (including those that has in Spain with Stellantis). According to Séjourné, “it is not right” that these companies are manufacturing their cars in Europe with Chinese components and Chinese employees, noting that their investment in creating a local network of suppliers is minimal. A good example is the CATL battery production plant in Aragón where it is expected that employ 2,000 Chinese employees. Photo | ANFAC and Renault In Xataka | Before opening its gigafactory, Zaragoza has a pending task: create a “chinatow” for 2,000 Chinese workers

What’s new and what improvements are there in the new version of the ChatGPT model with two personalities

by

Let’s tell you What’s new in GPT-5.1the new version of the model artificial intelligence of ChatGPT. GPT versions are the engine of your interactions with OpenAI’s AI, and the results that are given to you and the way in which they are told to you depend on them. This new update stands out above all for having two versions with different personalities. But in addition to this, there are also other new features that go more under the hood, but that can also make a difference when it comes to serving you the answers. Here, you must be clear that this new version of GPT-5.1 It has reached paying users firstwhether they have a Plus, Pro, Go or Business subscription. Maybe later it will also reach free users, but probably with limited use. Two GPT-5.1 with two ways to respond As we have told you, the main novelty of this new update is that it offers two versions with different personalities. This goes beyond customize ChatGPT with different personalities as you can do in the settings, but there are directly two versions of GPT-5.1, and each of them has a different type of response. On the one hand it is GPT-5.1 Instantwhich is more conversational and with more “warm” or close responses. and then you have GPT-5.1 Thinkingthat use clearer language with less jargon. This last model is trained for deep reasoning, and responds faster on simple tasks, while dedicating more thinking time to complex ones. Paid ChatGPT users will see the ChatGPT 5.1 model activated at the top. By clicking on the name the model selector will open, and in it you can choose between the variants instant and thinking. There is also a way Car which will choose for you which of the two variants to use depending on what you ask in the prompt. Smarter adaptive reasoning GPT-5.1 also improves its internal logic, and now dynamically decides the “thinking time” that you dedicate to a request. Come on, instead of dedicating the same time to each of them, you will dedicate different times depending on the type of request. This way, when you make a simple query it will be processed with minimal calculation and faster, while more advanced reasoning tasks receive additional layers of analysis to improve the results and make them more coherent and context-sensitive. Behavioral improvements and instructions This new model too improves following instructionsbetter “understanding” what you ask and generating responses that are more aligned with it. Each of the two variants adjusts its reasoning to the complexity of the request to ensure that the answers are consistent with everything you ask of them. Better tone and personality controls We already told you that ChatGPT has a setting to determine the tone and personality of the responses. Now, instead of having predefined ringtones the user can configure it. For example, you can choose to make it professional, friendly, or efficient, and apply it consistently across all interactions. For regular users this simply helps you be more comfortable with the answers it gives you. But for companies it is even more important, since you can align the tone with what you want to use both in your communications with customers and in internal documentation. Context retention improvements Context retention is more effective, which improves continuity in long interactions and with multiple shifts. This will help you as a user, but it is especially important in the business environment, in uses such as customer service or knowledge base systems. Performance optimization Response generation is now faster, and token overhead is reduced to make GPT-5.1 a better model for automated environments. It can deliver higher quality or better results using fewer tokens than previous models, reducing the overall costs of using the API. In Xataka Basics | The best prompts to save hours of work and do your tasks with ChatGPT, Gemini, Copilot or other artificial intelligence

what it is, characteristics of this artificial intelligence model and differences with Gemini and ChatGPT

by

Let’s explain to you what is Kimi K2 Thinkingthe latest model of the company artificial intelligence Kimi AI. It is an AI that has made a name for itself today due to its open nature and for having managed to compete directly against GPT-5, Gemini 2.5 Pro and other high-end models. We are going to start the article by explaining what Kimi K2 is and the characteristics that make this artificial intelligence model different. Then, we will finish by telling you the main differences with respect to the most popular models on the market. What is Kimi K2 Thinking Kimi K2 Thinking is the latest version of Kimi, a Chinese artificial intelligence model created by Moonshot companyfrom Alibaba. Since the names are the same, you can clarify yourself by thinking that Kimi is the company’s AI, like ChatGPTand that this AI has different models that are being launched, such as the case of GPT-5 in the case of OpenAI. Kimi K2 was launched in July, and stood out for having a gigantic size of 10,000 million parameters. Now there is a new version called Kimi K2 Thinkingwhose number of active parameters amounts to 32,000. According to its creators, this allows the AI ​​to maintain stable use of agentic tools over 200 to 300 sequential calls. And what does all this talk mean? As you know, we are entering the era of AI agentswhich are automations with which an artificial intelligence can carry out different actions autonomously. This allows the AI ​​to even make decisions for you, from asking it to make your purchase to preparing a vacation package for you and taking care of the reservations. It is also something that at the business level is going to have even more uses. Therefore, the more ability an AI has to perform a large number of actions without making mistakes, the more valuable and powerful it is. Features of Kimi K2 Thinking The most important feature of Kimi K2 Thinking is that it is an open model. The models of companies like OpenAI, Google or Anthropic are closed, which means that their source code is kept under lock and key, only these companies know how it works inside. Meanwhile, K2 Thinking is open source, which means that anyone can know how it works inside looking your Githubits features, and you can even adapt it for free. What’s more, you can install it locally at no cost, although the computer needed for this is too powerful for ordinary mortals, but “distilled” versions, lowered or trimmed, can be released so that people can use them locally. In this aspect it is like DeepSeek, another open AI that already surprised a few months ago for approaching the power of non-open models such as Gemini or ChatGPT. In the case of Kimi K2 Thinking, according to the test benches has managed to surpass GPT-5something that until recently was unthinkable. We are facing a Mixture-of-Experts architecture model (MoE), which means that it is made up of several experts (subnetworks or specialized modules), and that not everything is activated at once, but only the parts of the model necessary to answer what you ask or perform the task that you have asked. It should also be said that it is multilingual, and can be used in other languages ​​although it focuses on Chinese, and that it can process many types of file formats. Also searches in real time to offer you the most up-to-date information, and is multimodal, being able to interpret text, images, code or a combination of these. Kimi K2 Thinking can be used as a conversational char answering questions and maintaining long context while following complex threads. But it can also interpret images, or a combination of mixed inputs such as images with text and with code. In addition to this, it can generate programming code, analyze long documents thanks to its large context window and extract information to answer questions about the content or give you a summary. Additionally, you can create automations or agents. Differences with ChatGPT or Gemini As we have told you above, the main difference of Kimi is its open concept. While ChatGPT and Gemini are proprietary models, Kimi allows access to the community so they can see its code. Several benchmarks have shown that Kimi K2 Thinking outperforms GPT-5 and Claude Sonnet 4.5 (Thinking) in search and agentic browsing in the browser, in text-only operation, and in information collection. The only thing in which it still does not surpass these models is in the creation of code. In the use of agentic tools, benchmarks or test benches have shown that Kimi K2 Thinking is positioned as a leading AI model. Besides, Kimi is a cheaper model for several things. First, training the model cost $4.6 million, according to indicate on CNBC, a ridiculous figure considering that training proprietary models like GPT-5 It cost about 500 million dollars according to estimates. It’s also cheaper to use the Kimi K2 Thinking API. The API is like the entry key that allows other applications to connect to this AI to work with it. The price of K2 Thinking is $0.6 per million tokens in and $2.5 per million tokens out. GPT-5 Chat costs $1.25/10 respectively, and Claude Sonnet 4.5 costs $3/15 respectively. For the average user, the operation is the same.. You have the website kimi.comwhere after registering for free you can use the Kimi K1.5 and K2 models. However, If you want to use Kimi K2 Thinking you will have to pay with their subscriptions of 19 or 30 dollars. At least, this is if you want to use the full version on the official website, without having to install anything. In Xataka Basics | The best prompts to save hours of work and do your tasks with ChatGPT, Gemini, Copilot or other artificial intelligence

We believed that no open model could outperform GPT-5. A Chinese startup proves us wrong

by

A Chinese startup called Moonshot just launched Kimi K2 Thinkinga gigantic open model with a trillion parameters that has done something that seemed almost impossible: surpass the best proprietary models from companies like OpenAI, Google or Anthropic. If we thought that “Open Source” models could never compete with GPT-5, Gemini 2.5 Pro or Claude, we were wrong. what has happened. This “AI laboratory” had already announced Kimi K2 in July with that gigantic size of one trillion parameters, but now they have released the “Thinking” version with that same size (32 billion active parameters, Mixture of Experts architecture). According to those responsible, the model is capable of maintaining stable use of agentic tools over between 200 and 300 sequential calls. Or what is the same: it can chain long sequences of actions autonomously and apparently without error. The best of all is not that: it is that it surpasses GPT-5 or Claude Sonnet 4.5 in various tests and costs much less than those models. The benchmarks. Those responsible for Moonshot explained how Kimi K2 Thinking achieves the highest scores in Humanity’s Last Exam (general knowledge, 44.9%) and BrowserComp (agent browsers, 60.2%). He is almost at Claude’s level in the SWE software development test, and is also almost the best in another of those benchmarks, LiveCodeBench v6. It is true that in some tests still slightly behind of its “western” rivals, but the achievement is spectacular. More benchmarks. Those responsible for Artificial Analysis have shown their first conclusions after evaluating it with various tests. Thus, they highlight its behavior in agentic tasks that simulate that the model is acting as a customer service agent. In this test it obtained 93% of the maximum, surpassing all its competitors by far (GPT-5 Codex High obtained 87%, for example). They will do more tests, but for now the prospects are fantastic. And on top of that, cheap. On CNBC indicate that training the model cost $4.6 million, a ridiculous figure considering that training proprietary models like GPT-5 It cost about 500 million dollars according to estimates. Using the Kimi K2 Thinking API is also very affordable: $0.6 per million tokens in and $2.5 per million tokens out. GPT-5 Chat costs $1.25/10 respectively, while Claude Sonnet 4.5 costs $3/15 respectively. The details. The model makes use of an INT4 quantization to improve its efficiency without compromising the precision and quality of its responses. Its context window—the “size” of the data we can enter when making prompts—is 256k, a relatively modest figure for large models but still notable. And as a good open model, we can download it to use locally… if we have a real monster at our disposal. The model weighs 594 GB, and for example joining two Mac Studio M3 Ultra It is possible to make it work locally relatively smoothly at about 15 t/s. Alibaba is behindyes. Although the model is developed by an independent startup called Moonshot, this firm has been financially supported by Alibaba, which is becoming an absolute powerhouse in this field. Already not only conforms with developing its own models, which are outstanding (Qwen is the clear example), but is also financing the development of other models such as Kimi K2/Thinking. China and its love for open AI models. During the last few months we have seen how China dominated in the field of open AI models —not “Open Source”—. The Asian giant has adopted an overwhelming philosophy with increasingly better models but which until now seemed to be several steps behind the large proprietary models of OpenAI, Anthropic or Google. This is no longer the case. The race is lively. This achievement represents a new vote of confidence for the open models coming from Chinese companies. It is true that they are huge and that makes it very difficult to use them in practice by end users, but they present an interesting alternative for companies. Image | idnaklss with Midjourney In Xataka | There are many “internal” races within the greater AI race. And Alibaba is winning Open Source

Chinese hypermarkets are in crisis and have found the solution: follow the Mercadona model

by

The golden age of Chinese hypermarkets is coming to an end. With the economy stepping on the brake, these mastodons are in a tighten and desperately seek new formulas to hook consumers who look more at the pocket. In this new panorama, the solution seems to be betting on the strategy that Mercadona dominates perfectly for years. What’s happening. The great Chinese supermarkets are having You would be difficult to survive. In recent years, Carrefour has closed more than 140 stores, Tesco has disappeared and last year the main leading hypermarkets had Important losses. With The economy in decelerationChinese consumers are more cautious when spending and that is causing the main chains to change their strategy drastically, as reported in Bloomberg. The Mercadona model. Many neighborhood stores and more white brands, this is how some Chinese giants are adapting to this new era. The own brands were not usual in China, but currently they take more and more space in the halls of the main chains. In addition, they are beginning to change their store strategy, favoring the proximity of smaller stores instead of hypermarket that forces us to move by car and plan a larger purchase. Adapt or die. Chinese hypermarket chains are transforming with smaller formats and their own brands. Walmart, with its stores proximity to Lo Carrefour Express and its MarketSide brand, is a good example of this trend. The Wumart Group has launched Six stores with discounts in Beijing and FreeShyppo, from Alibaba, already has more than 300 stores under its cheap chaopa brand. Approximately 60% of the products found in these stores are white brands. This strategy responds to the search for savings and convenience by the consumer. The Pangdonglai case. It is a Henan supermarket that has achieved viral success. Its strategy is based on exceptional customer service, good treatment of unique employees and services such as ticket offices with dog water and personalized preparation of the purchase basket. But the main secret of their success is that they have placed their profit margin in 30%, which allows them to keep low prices all the time, without having to resort to specific promotions. Despite having been born in a smaller city, its model is so influential that Yonghui Superstores, the fourth chain of China, is reforming its stores following its example. Image | Wikipedia In Xataka | The US studied what would happen if it enters war with China. Now he has started a career desperate to double missiles

Spanish Clevergy has just lifted 3.2 million to expand its energy management model

by

The relationship between marketers and households is changing: it is no longer going on invoices, it goes from apps that explain what your home consumes and what you can optimize. In that day -to -day landing stands out Clevergya Spanish startup founded in 2022 that has just closed 3.2 million euros To make your European leap. Its proposal allows companies to offer not only a personalized application with real -time monitoring, alerts and savings recommendations, but also a set of solutions to digitize their business. The promised result: customers who better understand their energy consumption and companies that modernize their offer without starting from scratch. Founded in Madrid in 2022 by Beltrán Aznar, Álvaro Pérez and Juan LópezClevergy has moved quickly in a sector where digitalization is already a demand. In just three years he has managed to arrive, according to the company, “hundreds of thousands” of Spanish homes through their agreements with marketers. Its role is clear: it acts under a B2B2C model, that is, it offers technology to companies so that it is the ones that put it in their end customers. This combination of speed and adoption has given visibility in a market in full transformation. Clevergy seeks to convert energy management into a daily experience Clevergy’s proposal for marketers goes beyond an app for its customers. The company has developed a portal that allows to centralize operations and support, in addition to identifying business opportunities and cutting costs. It also offers one API to integrate consumption data and generation from counters, solar panels or connected devices. To this are added white brand applications, adaptable to the identity of each company, and modules that can be inserted into existing platforms. For homes, all this deployment is concretized in functions designed to give more visibility about the energy they consume. Customers can monitor their real -time spending, receive notifications when inefficiencies are detected and adjust their consumption habits. The system also includes comparisons with other users, calculation of potential savings and remote control of connected equipment. In this way, marketers seek to add a tangible value to their offer and generate confidence in a market where the price is no longer the only decisive factor. Clevergy’s growth has been fast. In just three years he claims to have tripled its growth and, in just 18 months, has closed two rounds of financing: the first of 1.5 million euros in 2024 And the second, of 3.2 million, is the one that has just been announced in 2025. The latter is the one that marks a turning point, when arriving at a time when marketers intensify the search for digital services to improve their relationship with customers and reduce costs. For the company, it is a validation of its role in this transformation process. Clevergy has closed two rounds of financing: the first of 1.5 million euros in 2024 and the second, of 3.2 million The round of 3.2 million euros has been led by Racine2 (managed by Serena and Makesense) together with Axon Partners Group, with the participation of Satgana, Wayra (the CVC of Telefónica) and Angels, Juan Roig’s investment society. With these funds, Clevergy seeks to accelerate its international expansion and improve the capabilities of its platform. The declared objective of the company is to continue refining its technology and progressively take it to other countries of the continent. The challenge is now to check how far Clevergy can go outside of Spain. The company has shown traction in the national market, but the jump to Europe implies integrating with different regulations and compete in a stage with other technological and energy actors. It will be key to see how it manages to deploy its platform in new countries and if the marketers really transfer that proposal to the final customer. Its evolution will mark to what extent this digitalization model can be consolidated beyond the domestic market. Images | Clevergy In Xataka | Juan Roig believes that in the future no one will have cooking at home. Mercadona is conquering the market thanks to it In Xataka | A Basque startup of AI has just lifted 189 million euros with a great idea: compress the AI

Log In

Forgot password?

Forgot password?

Enter your account data and we will send you a link to reset your password.

Your password reset link appears to be invalid or expired.

Log in

Privacy Policy

Add to Collection

No Collections

Here you'll find all collections you've created before.