Nobody would bet on Meta today in the superintelligence race, but he plays with more advantage than we think to win it

A year ago, Zuckerberg was hiring AI talent like the world was ending, offering millionaire salaries and even buying entire companies to be able to sign Alexandr Wang. One year later, 8,000 people have been laid off, the work environment is unbreathable and We are still waiting for them to launch that great model with which to compete with OpenAI and Anthropic, all this while they spend money a lot. Despite everything, Meta has a real chance of closing positions and getting closer to the podium in the AI ​​race. The near future. In a complete report by Semianalysis They talk about how Meta is playing with better cards than it may seem. Muse Spark, its first language model, was somewhat disappointing, falling behind Chinese competitors such as Deepseek v4 Pro or Kimi K2.6. But the important thing is not where Meta is now, but where it can be in the near future thanks to the combination of three key elements: data, talent and computing. Record employees. It was a very controversial decision and, as expected, Meta employees were not amused. Without them being able to object, software was installed on the company computers that I recorded everything they didnot to spy on them, but to train their AI. This data is pure gold for training agents: Meta is accumulating thousands of examples of different people solving the same tasks, while data companies like Surge or Mercor have to partner with others to be able to record their workflows. Meta has the data at home. They say in Semianalysis that this decision is as if they had created a “top-tier startup for RL environments” within the company, with one of the founders of Scale AI leading the transformation. In addition, after the restructuring they have put at least 3,000 engineers on reinforcement learning environment tasks. All this data is key to being able to create programming agents like Claude Code or Codex from OpenAI. Data centers. It is one of the main sources of spending for Meta, which is building several gigantic data centers whose capacities are more than 1 gigawatt. Maybe Meta cannot compete in infrastructure with hyperscalers like GoogleMicrosoft and Amazon, but things change if we confront it with frontier AI laboratories. Here, Meta has a clear advantage and, according to Semianalysis’ projections, Meta will have more computing power than Anthropic and OpenAI combined before the end of the year. The talent. Last summer, Meta began signing talent with a checkbook. They hired at least 14 high-level researchers who came directly from Anthropic, Google and OpenAI, paid $14 billion to keep Alexandr Wang and Scale AI. Bringing together the best does not ensure that the team will work and, in fact, for months now there have been rumors of internal tensions. Of course, if they make it work, they have the talent. keep focus. Meta may be in the rearview mirror of OpenAI and Anthropic sooner rather than later, but it is one thing to have the resources and quite another to achieve it. Meta is in a delicate moment internally, with many employees very dissatisfied with the company’s strategy. If they do not navigate these waves well, they risk becoming unfocused and lost along the way. Image | Xataka with Magnific In Xataka | Meta has a long history of privacy scandals. We can add one more to the list

the biggest bet in its history to national chip manufacturers

China has defined a series of steps to become the first world power. Technological sovereignty falls within this roadmap and, to achieve the objective, a greater investment in local technology is necessary. The current battlefield is that of AI chips, a segment that Nvidia dominates with an iron fist, but if there is a way to change the trend it is with technology and money, a lot of money. And China has found its three winning horses. In short. China is a giant market for Nvidia. With Western Big Tech in his pocket (although Anthropic and Meta are looking favorably at Samsung for inference chips for the age of agentic AI), Nvidia was missing the other big pie. China is a market valued in 50,000 million dollars for whoever manages to position itself as a leader in the AI ​​chip segment, but the problem for Nvidia is that it is not easy at all. Due to bureaucratic problems on the part of the US and China itself, Nvidia is a bit in no man’s land. Meanwhile, national companies have been positioning themselves, presenting their alternatives and GPUs for training this technology. And with the order to bet on the national product, Chinese companies are clear about it. As we read in Business TimesChinese Big Tech currently invests 30% annually of its budget in national AI accelerators, but that spending will increase to 46% over the next twelve months. The horses. Some of these Big Techs are creating their own chips, or already have them. This is the case of Alibaba or Tencent, which would be a kind of Meta or Google: they offer a product to the user, to other companies and, in addition, they have their own data centers with their home-made chips. However, there are other companies that are more focused on creating the hardware to power the internal data centers of those other large companies. This is the case of Hygon Information Technology or Cambricon Technologies, but also of Huawei. All three offer AI platforms that are becoming increasingly important in a context in which Nvidia H20 chips designed specifically for China They are increasingly difficult to find due to that national order to prioritize homemade hardware. The advantage of these companies is that it is no longer about raw power, which is one of the advantages that Nvidia had (although Nvidia is much more, since it also has a complete software suite and so on), but about speed and low latency. For agentic AI, high bandwidth for fast data transfer is essential, and that seems to be a point that Chinese hardware companies are focusing on. The Ascend. The interesting thing is that this strategy is not a future plan and is already being put into practice. Zhipu AI, for example, is a company that has already trained its GLM-5 model on more than 740 billion parameters entirely with the Huawei Ascend 910C. In addition, the Chinese company is working on the development of the Ascend 950 and 960 chips that promise to achieve parity with Nvidia’s most modern architectures. And the supernodes. On the other hand, Huawei continues to push the Atlas 950 supernode. Presented a few months ago, it is a cluster of up to 9,192 Ascend 950DT NPUs per system with up to 1,152 TB of unified memory. It is, according to the company, a direct competitor to Nvidia’s NVL144 rack, achieving 6.7 times more computing power than that of the American company. It is designed for both training and inference and, although it raises doubts about the energy efficiency per computing unit compared to that of Nvidia, what is clear is that it is a product that Chinese Big Tech They will be able to ‘shoot’ for their future models. Huawei is taking the platform around the world and its next appointment is at the World Artificial Intelligence Conference 2026 in Shanghai. Whether Nvidia solutions are used or not, what is clear is that the intention is to prioritize national hardware. China is going to invest nearly $300 billion to build data centers over the next five years to boost AI across all sectors of society. For its part, at least 80% of the basic technologywhere the chips come in, will be supplied by national companies. As several have already pointed out (Jensen Huang among them), the US veto of China It was the worst move they could make. because they only achieved one thing: give wings to their accelerated technological development. In Xataka | Huawei is very close to a great technological milestone. And he is very clear about what has allowed him to do so: the US veto

Microsoft bet 80 billion that you would play hundreds of titles. It failed because you just wanted to play the same old game.

I have dozens of games in my library. Do you know how many I play? to two: from time to time I play a FIFA (I’ll always continue to call it that, I’m afraid) and if I get around to it I’ll hit the old ‘Battlefield 1’ that still has me conquered. In recent years I have enjoyed ‘Ghost of Tushima’, ‘Star Wars Jedi: Survivor’ or ‘Sifu’, for example, but while I was spending hours on those half a dozen titles, many others that I have in my library are simply collecting dust. Surely many of you play many more games and invest much more time in this, and Microsoft was surely thinking of you when it configured its future strategy. Xbox’s ambition was great until it wasn’t The idea in 2020 was clear, and those responsible for Microsoft said that There were 2.8 billion gamers in the world and they wanted to conquer them all. As? Well, with two great projects: Spend a million in buying video game development studios to constantly grow the catalog, and, of course, Create the ‘Netflix of video games’ with Xbox Game Pass Both ideas seemed to make a lot of sense and many gamers, especially those of us who were from Xbox, felt excited by that ambition. Xbox Game Pass seemed to be really exceptional for gamers and a nightmare for Sony, and although the operation to buy Activision Blizzard and franchises like ‘Call of Duty’ It was exceptionally expensive.the play again seemed like a fantastic bet for players of all types. But lo and behold, it wasn’t. These days we have known that Microsoft has spent close to $80 billion on agreements that would give it access to titles such as the aforementioned ‘Call of Duty’ or ‘Skyrim’, so that these video games could be available from day 1 or almost from the first moment on Xbox Game Pass. They wanted to create a catalog that would allow gamers to play hundreds of titles, including the most recent “triple A” titles, as part of the subscription to that service. They were wrong. And we also thought that the idea made sense. less is more The truth is that the strategy has not worked out well for them. The company has announced that will lay off 3,200 employees in its Xbox division20% of its staff, and will also allow five development studios to leave the company to follow their own path again. Maybe the answer wasn’t having more games. Steam already gave a strong clue at the beginning of 2025 when it pointed out that its platform It had become a video game dump.: Of 19,000 titles in 2024, almost no one played 80% of them. As in many other areas, the long tail theory prevails: interest is concentrated in a handful of contents that, on their own merits (perhaps seasoned with some luck) become absolute successes. The reality for the vast majority of gamers is the same: there are too many games on the market and people simply don’t have time to play all of them. Let’s be handsome or uglyWe insist, there is no time for so many video games. My colleague Jose García wrote in 2021 his particular reflection on how many we have become new digital Diogenes When it comes to video games: we get countless new games (most of them free) on Epic Games, Steam or GOG, but we have them collecting virtual dust in our library because we will never play many of them. My situation was also reflected in another text by Isra Fernández: Being a gamer after having children is complicated. The Netflix of video games was a utopia The picture that this situation paints for us is one that we intuited and made sense, but that we deceived ourselves by rejecting. Creating the Netflix of video games was impossible because video games are radically different to series or movies. While one consumes series or movies once and abandons them probably forever, gamers return to the video games we like again and again. We don’t spend an hour or two on them, but dozens and even hundreds of hours both in solo games and in games against other players on the internet. The way of consuming both content is precisely what destroys the illusion that a Netflix of video games makes sense. Xbox Game Pass offers us hundreds of titles so we can try them (and buy them) if we want, but we rarely do. I myself enjoyed Xbox Game Pass Ultimate for three years, and although I tried around twenty titles during that time, I always ended up returning to my classics. The options were wonderful, but I just didn’t have time for them. because when I wanted to play, I wanted to go back to my FIFA or my ‘Battlefield 1’. Microsoft itself already accused the problem when it gave us the worst news two years ago by raising the price of Xbox Game Pass. That was not enough, because a few months later he uploaded it again. The service was becoming more exclusive than ever, and it was doing so for one simple reason: they weren’t making enough money with it. But there is another problem that It is not just about consuming, but about producing. The “Netflix rhythm” is also unviable: there four scriptwriters start writing and in six months the platform releases a series that people devour in a weekend. In video games things are different, with increasingly longer and obscenely expensive developments. The famous availability of games available on “day 1” does not help either. That AAA video games worth 70 or 80 euros are available for free to subscribers was wonderful for us, but a shot in the foot for the company’s accounts. Building loyalty was important, but that was killing income for developments like these in which a lot of time and money is invested. The conclusion is clear in retrospect: Microsoft’s strategy was fantastic for … Read more

bet on an NVIDIA ARM chip

Microsoft has been looking for a convincing answer for a long time to a question that the PC has been dragging on for years: whether Windows on ARM can be something more than autonomy, silence and light productivity. He Surface Laptop Ultra It is born precisely there. The company describes it as its most powerful Surface Laptop yet and ties it to more ambitious workloads, from creation to local AI, leaning on NVIDIA RTX Spark as a great novelty. Microsoft’s move comes after a move that did not go unnoticed. Apple announced in 2020 the transition from the Mac to Apple silicon and materialized it that same year with the M1the company’s first own chip for its computers. With it, performance per watt, unified memory and vertical integration became an advantage that was difficult to ignore. That pressure no longer comes only from the high range. He MacBook Neowith a starting price of 699 euros, came to compete with affordable Chromebooks and Windows laptops. In parallel, Microsoft has tried to organize its own response with the Copilot+ PCa more solid category than its previous attempts with ARMbut still far from becoming a bet that has paid off in popularity and, consequently, in market share. A laptop to demonstrate an idea: Windows on ARM wants to go beyond autonomy The file that Microsoft has advanced It has two layers. The first is in performance: NVIDIA Blackwell RTX GPU, up to 128 GB of unified memory, CUDA and 1 petaflop of AI compute (based on theoretical FP4 TOPS with sparsity). The second is in everything you see and touch: a touch screen mini-LED 15-inch PixelSense Ultra, up to 2,000 nits of brightness HDR maximum, 262 pixels per inch and the largest haptic touchpad that Microsoft has included in a Surface. There is also something almost surprising in 2026: HDMI, USB-C, USB-A, SD and headphone ports without depending on adapters. When Microsoft talks about unified memory it doesn’t mean putting a lot of RAM in a laptop. The important thing is that this memory can serve as a common room for CPU and GPUso that the workloads are distributed according to what each task needs, although it eliminates the possibility of later extensions. The other half of the announcement is in what has not yet appeared. Microsoft has not detailed the price, the exact launch date, the final configurations, the specific processor, the exact GPU or the autonomy expressed in hours. One detail: the company itself remembers in the fine print that the Surface Laptop Ultra is a preliminary product, that its characteristics may change and that the sale will depend on the certifications or regulatory approvals of each country or region. Microsoft doesn’t get to this point from scratch either. The Copilot+ PCs of 2024 had already served to organize a new stage of Windows on ARM, with processors capable of surpassing the 40 TOPS of NPUlocal AI functions and equipment from partners such as Acer, ASUS, Dell, HP, Lenovo and Samsung, in addition to the Surface themselves. However, the available data invites us to put the growth in context. The Registerciting Context, placed Copilot+ PCs at only 9% of AI PCs distributed in Europe in the second quarter of 2025. The figure showed progress, but also a still small base. That’s where NVIDIA changes the type of conversation. The first wave of Windows on ARM relied heavily on efficiency, NPU and autonomy, but the Surface Laptop Ultra wants to add an ingredient that was missing from that story: graphical muscle and compatibility with a well-established professional ecosystem. CUDA It is not a minor word for anyone who works with AI, rendering or professional software supported by the NVIDIA ecosystem, because it has been part of many workflows for years. Microsoft is also introducing optimized Prism, native creative app support for ARM, and agreements with partners in gaming and anti-cheat systems. What we have before us seems like a very interesting proposal. The announcement invites us to take it seriously, because the Redmond firm has made its position quite clear: it wants this Surface to be proof that this new stage of Windows on ARM can go much further. But it is still incomplete to pass sentence.to. We’ll have to get our hands on it to see how it performs, how it behaves and how it works in real life. Images | Microsoft In Xataka | Apple has shown the weaknesses of laptops with the MacBook Neo. And the manufacturers are getting their act together

If Europe wants to bet on its agriculture, it is doing it wrong

In the last decade, the European Union has lost three million agricultural holdings. That is almost 25% of all those that existed in 2013. And it is curious because, in its last cycle alone, the Common Agricultural Policy is spending 387,000 million euros. What are we spending all that money on? A poorly posed question. I recognize it. When one sees the enormous amount of money that the CAP moves and its central role in the Union’s debates, one tends to assume that this “silent” transformation of the European field is taking place. despite of Brussels policies. However, when we look at the data, doubts arise. According to the European Court of Auditors and the Commission20% of the beneficiaries take around 80% of the funds. To the extent that the CAP distributes direct payments based on hectares, we can say that those 20% are the large landowners. In Spain, for landing the datathe 1% that receive the most help concentrate 28%; 10%, 62%; and 20%, 79%. It is a scheme that strongly encourages concentration. And it shows. Continuing with the Spanish case, the country has lost 180,000 farms in the last 15 years. Only between 2020 and 2023, 12.4% were lost. In this he has had a lot to see about the pandemic and the Ukrainian Warbut everything takes on a new prism when we realize that macro agricultural holdings grew by 6%. The result is that that 20% that accumulates 79% of the CAP represents 7% of the agricultural area. We are seeing it in the crown jewel of the Spanish countryside: the olive grove. As Datadista and Greenpeace explained, the accelerated entry of investment funds has disrupted the sea of ​​olive trees. We are moving from a traditional dry-land olive grove to a super-intensive hedgerow olive grove on irrigated land. But… is this a problem? To answer this question, it is best to look at the Netherlands, the European “crystallization” of what is at the end of the trend that drives the CAP. Netherlands is second (or third, depending on the year) agri-food exporter in the world and it is in a territory smaller than Galicia, with 1.81 million agricultural hectares and only 52,106 farms. That is to say, its productivity per hectare is crazy. For this reason, the concentration of the CAP is also extreme: 1% takes 40% of the funds. The problem is that it is not environmentally sustainable: between 2022 and 2024, nitrogen emissions caused an unprecedented political crisis. Something that, with the slurry ponds either the crises of the Mar Menorwe are also seeing in Spain constantly. And now what? That is the central question. Because this CAP lasts only until 2027 and just now we are starting to discuss what we want for the future. Considering how quickly things have changed in recent years, that future may be completely different from anything we know. Image | Rob Mulder In Xataka | In California, the funds discovered that there is no investment more profitable than farmland. Now it’s Spain’s turn

Telecinco faces one of the toughest summers in its history. He has bet all his chips on Paz Padilla

Last August, Telecinco closed the month with an 8% audience share, the worst monthly figure in its history. In May 2026 it averages 9.2%: somewhat better, but it certainly does not provide a much better diagnosis of the state of suspended animation that the channel lives. Mediaset has been setting annual historical lows for four consecutive years, and the response for this summer has a name and surname: old acquaintances such as Carlos Lozano or Paz Padilla return to the channel, in a decision that makes it clear that Mediaset’s bet is not based on risk. Four years in free fall. The channel signed its historic low audience for the fourth consecutive year in 2025, with a 9.5% annual feeafter 2024 had already been the first below 10%. The problem is not only with the public, but with everything that comes with it: the net advertising income of Mediaset España fell 8% in 2025. For the first time, the group was surpassed in advertising billing by Atresmedia. To this we can add a structural crisis of the environment: Daily linear television consumption time fell by 7.7% in 2025, with an especially pronounced contraction among viewers aged 14 to 49. Telecinco competes, like all of them, against the changing habits of a generation more attentive to streamingTikTok and video games. Lozano, Padilla and other old acquaintances. Under the motto “Open for holidays”, Telecinco has presented its summer strategywhich includes the dating show ‘Love or whatever arises’ (with Carlos Lozano returning to the channel eight years after leaving it, reviving the spirit of ‘Women and men and vice versa’ in the same afternoon slot) or the daily version of ‘De friday’ called ‘Monday to Friday’with Santi Acosta and Beatriz Archidona. Finally, the entertainment format ‘The Peace Show’ with Paz Padilla at the helm, the weekend afternoons and reducing the ‘Party’ time. The presenter was fired by the channel in 2021 in quite tense circumstances, which says a lot about how both parties need each other, in a cessation of hostilities pact that must not have been easy. The changes will start in the second half of June. What goes away and what resists. To adjust this whole grid, there are programs that go on vacation. ‘Jorge’s Diary’ takes a break until September, coinciding with Jorge Javier Vázquez’s vacation. Hold ‘Allá tú’ with Juanra Bonet in the afternoons from Monday to Friday thanks to its good audience of close to 10% of share. The same thing happens with ‘The Fair Price’ by Carlos Sobera, which will also try out on the weekends, filling the gap left by the resounding failure of ‘Seen what has been seen‘. The goldmine of dating programs. Along with the realitiesand sometimes incurring almost unnatural matings between both variants, the dating shows They are the only thing that now invariably works for Telecinco. ‘The island of temptations‘has maintained competitive audiences during this four-year crisis; ‘Married at first sight’ was last season’s surprise. ‘Love or whatever arises’ tries to get away from this trend, and also with a specific format that has been tested previously with success. However, even if all these proposals work, they will be nothing more than mere patches to liven up the summer of a chain that needs a deeper reinvention. The return of Lozano or Padilla are symptomatic of the lack of originality of the resources of a Telecinco that survives on the basis of patches and proper names that drag audiences down. Because an ‘Island of Temptations’ is something that only happens from time to time: to get there you have to keep taking risks and, possibly, try formats that don’t smell like you’ve already seen them. In Xataka | The last bullet that Telecinco has left in the audience’s gun is a promo generated by AI and based on a TikTok success

The bet of “Everything on subscriptions” has not turned out as Microsoft expected

Microsoft bought Activision for almost 70,000 million with a clear and undisguised idea: put the successive installments of ‘Call of Duty’ in Game Pass from day one. A strategy that, on paper, would boost the service’s subscribers and change the tides of the industry. Eighteen months later, perhaps the numbers have not turned out to be so favorable and the company is rectifying that strategy at full speed. The most curious thing? After some changes, users end up paying more for a service that offers less. Price changes. Six months ago, Game Pass Ultimate cost 17.99 euros per month and included from day 1 the latest ‘Call of Duty’ to date, ‘Black Ops 6’. In October 2025 Microsoft raised the price to 26.99 euros, 50% hitjust two weeks before the premiere of ‘Black Ops 7’. Now on April 21, 2026 announces which lowers it to 20.99, but without the future ‘Call of Duty’ on day one. PC Game Pass, in parallel, goes to 12.99 euros from the previous 14.99, although it is also above the 11.99 it cost before the last increase in price. Not so discount. Despite the apparent discount, the April subscriber pays three euros more than the subscriber in September of last year, but also renounces the main claim for the service. Future ‘Call of Duty’ will arrive on Ultimate and PC Game Pass “during the following holiday season, approximately one year after” their commercial release, while ‘Black Ops 6’ and ‘Black Ops 7’ remain in the catalog. Only half a year has passed after the total restructuring with the purchase of Activision… but with the rate somewhat more expensive than it was not too long ago. Why it didn’t work. The strategy made sense: if ‘Call of Duty’ is the most profitable franchise on the market, offering it on the first day within the subscription would make Game Pass an almost impossible proposition to refuse. The numbers did not match. A report last October estimated that Microsoft had missed out on $300 million in revenue by including ‘Black Ops 6’ on the service in 2024. That same report noted that 82% of the game’s full-price sales occurred on PlayStation 5. Things got worse with ‘Black Ops 7’. Due to its presence on Game Pass from day one, the game’s launch sales fell more than 60% in some marketsand in the United States it ended 2025 as the fifth best-selling title of the year, the lowest position for a franchise game in almost two decades. Subscription was cannibalizing sales without growing enough to make up for it. The accounts don’t work out. Perhaps Microsoft’s accounts collided with an indisputable reality: there is not enough Xbox to support the expenses of a blockbuster of the caliber of Call of Duty, mainly with subscriptions. In November 2025 Calculations placed Xbox Series X/S at 34.10 million units sold compared to 86.12 million for PS5. Of course, the difference grows quarter by quarter. Giving away a game on day one that still costs $69.99 on PlayStation meant giving up margin in the most profitable territory to monetize Call of Duty. What point are we at? Christopher Dring, editor of The Game Business, pointed that the decision has also been made due to an imminent launch: ‘Forza Horizon 6’ arrives on Xbox in May and is, right now, the third most desired game on Steam, with nearly 2.7 million people who have included it in their wishlists. It is a good asset, with its previous arrival on the console, to increase the subscriber base, and the price drop may interest more than one player with doubts. In Xataka | Game Pass is already an unsustainable investment: more than 2,000 euros for each generation of console and without anything owned

Your bet in the AI ​​race is to bring together several functions in a single model

The artificial intelligence race is often told as a competition to see who builds the most powerful model or the one that dominates the most benchmarks. In the middle of that board, the French startup Mistral AI has just presented Mistral Small 4a proposal that tries to occupy a different place in that conversation. It is not presented as a model limited to a single function, but as one that, according to the company, seeks to bring together several advanced capabilities within the same tool. What exactly is Small 4. The company presents it as the new great iteration of its Mistral Small family and, above all, as the first model of the house that brings together capabilities that were previously distributed among several lines. Specifically, it integrates functions associated with Magistral, Pixtral and Devstral along with those of the Small series itself. Fewer models, more features. One of the central ideas of the announcement is to concentrate tasks that are normally solved with different tools in a single system. According to Mistral, the goal is that the same model can be used to converse, analyze complex information, work with images or assist in programming without having to switch between several specialized systems. The numbers behind Small 4. The model is based on a Mixture of Experts architecture, a design that distributes processing between different specialized submodels and that today appears in several artificial intelligence systems. In the case of Small 4, Mistral indicates that the system has 128 experts and that only four participate in each generated token. According to the company, the model reaches 119B total parameters, with 6B assets per token, and offers a context window of up to 256k. Who is this model intended for?. Beyond its architecture, Mistral also describes quite clearly the scenarios in which it imagines the use of Small 4. Let’s see. Developers: Automate programming tasks, explore code bases, and code agent workflows Businesses: conversational assistants, document understanding and multimodal analysis Research: mathematics, complex analysis and reasoning tasks The underlying idea is that the model can move between quite different needs without forcing you to change the system depending on the type of work. The graphics. In the material accompanying the announcement, Mistral includes several graphs where it compares Small 4 with other models in different benchmarks. These comparisons are not limited to the score obtained in each test. They also show the average length of the responses each system generates, a data the company uses to illustrate how much text each model needs to produce to achieve certain results. One of the graphs in the advertisement corresponds to the AA LCR benchmark, where Mistral compares the scores of various models and the average length of the responses they generate to solve the same tasks. The data published by the company are the following: • Mistral Small 4: 0.72 score with 1,600 characters• GPT-OSS 120B: 0.51 with 2,500 characters• Claude Haiku: 0.80 with 2,700 characters• Qwen3-next 80B: 0.75 with 5,800 characters• Qwen3.5 122B: 0.84 with 5,700 characters The comparison. Small 4 is not the highest scoring model. Both Claude Haiku and the Qwen models appear higher in that indicator. However, Mistral highlights another aspect of the comparison: the length of the responses. According to the company, its model achieves this combination of score and output length by generating significantly less text than several of its competitors, something it relates to lower latency and lower inference cost. The short answer trick. A shorter answer is not better simply because it takes up less space. It is only if it manages to solve the task with a level of quality comparable to that of a longer answer. This is where Mistral tries to put the focus: if a model achieves a competitive result by generating less text, it can respond faster, consume fewer resources and reduce the cost of inference. In other words, the advantage is not in being more concise, but in needing less output to reach a useful result. How to access the new model. Small 4 can not only be used via API and AI Studio. Being published under license Apache 2.0is also proposed as an open model that can be downloaded, adjusted and deployed in your own environments. The company adds that it can be tried for free at build.nvidia.com, in addition to offering it for production as NVIDIA NIM. Images | Mistral In Xataka | OpenAI has been wanting to be the bride at the wedding and the dead man at the funeral for years: now it has finally defined its priority

bet everything on AI

Samsung Galaxy S26 Ultra, when the hardware is practically unbeatable there is only one plan left: bet everything on AI S-Day has arrived. Samsung has just made the Samsung Galaxy S26, Galaxy S26+ and Galaxy S26 Ultra official, its three new high-end proposals. A continuous evolution with some functions that we had not seen to date and a complete commitment to photography. We review the technical specifications and features of the Samsung Galaxy S26 Ultra, the new reference in the high-end Android and a real commitment to artificial intelligence. Samsung Galaxy S26 Ultra, technical sheet with specifications and price Samsung galaxy s26 ultra dimensions and weight 163.6 x 78.1 x 7.9mm 214g screen 6.9 inches Quad HD+ resolution AMOLED LTPO 120Hz Gorilla Glass Armor 2 processor Qualcomm Snapdragon 8 Elite Gen 5 for Galaxy Storage 12 + 256 GB 12 + 512GB 16GB + 1TB rear cameras 200 MP, f/1.4, OIS 50 MP, f/3.4, OIS, x5 zoom 10 MP, f/2.4, OIS, x3 zoom 50 MP, f/1.9 ultra wide angle front camera 12MP, f/2.2 battery 5,000mAh 65W fast charging 65W wireless charging operating system Android 16 based on One UI 8.5 connectivity 5G (2xNano + eSIM)Wi-Fi 7Bluetooth 6GPSNFCUWBUSB type C others IP68 Integrated S-Pen Samsung Dex price From 1,449 euros This is how Samsung makes money: the secret is in the IPHONE The return to aluminum and a very special screen The design of the Samsung Galaxy S26 Ultra makes it the thinnest Ultra yet. Despite this, the cooling system has been improved using a vapor chamber, which is essential to keep Qualcomm’s best processor to date at bay, which we will talk about later. We return to aluminum and say goodbye to titanium. A small leap back in the hardness of the material to achieve a somewhat less heavy design and, above all, be able to dissipate heat better. If we turn it around, we find a 6.9-inch screen, with QHD+ resolution and Samsung’s own technologies (ProScaler) to improve image scaling. This allows you to obtain extra sharpness in texts and deeper colors without giving up realism. But what is most striking about the panel is not its specifications: it is its privacy function. A function that is difficult to describe beyond “genius”, and one that the rest of the manufacturers will undoubtedly have to take into account so as not to be left behind. What are you looking at, gossip? The Samsung Galaxy S26 Ultra debuts the Privacy Screen function, one of the biggest advances we have ever seen in an OLED panel. the phone is capable of independently controlling the pixel emission in certain areas of the screen, so that they disperse light in a very specific way. Through this light dispersion system, the user who is using their mobile phone will be able to see all the contents, but if someone tries to gossip about us from the side, it will be impossible. It is the natural replacement for privacy screen protectors, which darken as the viewing angle of the screen varies. This function allows you to hide sensitive elements such as pop-up notifications, passwords, gallery, etc. How does it work on a technical level? In normal mode, the pixel emission is traditional. In security mode, the largest pixels in specific areas are turned off, leaving only those that emit light at a narrower angle active. In other words, Only the pixels that we can see from the front emit light. Let’s talk about chicha… and AI The Samsung Galaxy S26 Ultra is the only member of its family to incorporate the Qualcomm Snapdragon Elite 5the latest chip from the Taiwanese giant. It does so accompanied by up to 16 GB of RAM and 1 TB of storage, with the classic base configuration of 256 + 12 GB. Since it has the best possible muscle, the key is in AI. This year, Samsung wants the user to choose their own path. Gemini remains the heart of your AI ecosystem, embedded in features like Circle to Searchreservation management, natural consultations, etc. Perplexity It has a more agentic role, aimed at obtaining information from the system through natural conversation. Bixby, which is still alive, is integrated into practically all native applications, and allows you to interact with them using voice commands. Regarding new features, in addition to the one related to screen security, SPAM calls or calls not registered in the agenda now They will be attended by an AI agent which will notify us in real time about whether we want to pick up the call or not. This phone comes with Android 16 along with One UI 8.5a new version focused primarily on natural language with Bixby on both desktop and web search. The cameras At the hardware level, the Samsung Galaxy S26 Ultra repeats with a 200 megapixel main sensornow being 47% brighter compared to the previous generation. The telephoto lens increases resolution to 50 megapixelswith five optical magnifications and up to 10x hybrid. There are no changes in the telephoto lens with three optical magnifications or in the ultra wide angle lens, the two most accessory sensors in this model. But it’s not all hardware, Samsung has made some tweaks to improve this configuration. There are improvements to the noise reduction engine. The algorithm adapts to the lens you are shooting to apply the necessary settings. Super stabilized video has improved. New codec to record in 8K in a more compact format and without losing information. Through AI, we can add content to images using prompts. In our in-depth analysis we will check if the changes in the algorithm have been enough for this camera to aspire to the podium of the best photography phones of 2026. An increasingly complex and fierce competition. Versions and price of the Samsung Galaxy S26 Ultra The Samsung Galaxy S26 Ultra starts at a price of 1,449 euros, rising to 1,699 euros in its 512 GB version. It can be purchased on the official Samsung website and its … Read more

Kia needed an electric Sportage on the market. The Kia EV5 is an (almost) perfect bet for the European family

Kia has been building one of the most interesting ranges of electric cars on the market for years. The EV family has managed to establish itself as one of the most attractive and risky options. From the Kia EV6 and its particular design to the most rational EV3 and the monstrous EV9. Now, the company has placed the EV5 on the market, one of the most rational proposals and necessary for your current offer. South Koreans needed a car that would perform the functions of the Kia Sportage, one of their best-selling models, with completely electric technology. And his proposal is as solvent as it is rational and attractive. Kia EV5 technical sheet New Kia EV5 Body type five-seater SUV Measurements and weight 4,610 meters long, 1,875 meters wide and 1,680 meters high. Wheelbase of 2,750 meters. 1800 kg weight. Trunk 566 liters with the sum of the front and rear trunk. Maximum power 160 kW (217 HP) and 295 Nm. WLTP consumption 16.9 kWh/100 km DGT environmental distinctive Zero emissions. Driving aids (ADAS) Mandatory by the European Union. Others Triple screen: 12.3-inch instrument panel 12.3-inch central screen 5-inch climate control screen Android Auto and Apple Car Play compatibility. Wireless mobile phone charging. Harman Kardon sound system as option. Electric hybrid. No. Plug-in hybrid. No. Electric Yeah. 81.4 kWh battery with 530 km of WLTP autonomy Versions with double motor (all-wheel drive) and a more powerful GT option will arrive. Price and release Now available With 81.4 kWh battery from 46,070 euros before aid (from 39,490 euros with discounts and aid) Why does an electric car have less autonomy than advertised? Balance is the word We could say explain the Kia EV5 with a football simile. The Kia EV5 is like a sober doorman. If you don’t like football, a goalkeeper sober He is the one who flees from eccentricities, the one who turns spectacular saves into simple saves. And a stop is just the final result of a very in-depth previous exercise, of strenuous training to be strong in the legs and extensive knowledge to position oneself in the right place at the right time. Whether the stop is complicated because it is attached to the lower corner of one of the posts or to give security to the team by taking the ball in a lateral center. Can an eccentric goalkeeper be good? Yes. And very good indeed. There are goalkeepers who earn their fame for stops that seem impossible, for having reflexes typical of the animal world. But it is no less true that many of these saves are only the result of having made a bad previous decision, of reaching the ball in a hurry for the simple fact of being worse positioned under the goal. Something like this happens with the Kia EV5. It is not a spectacular car in any sense. But almost everything is done grating at a very high level. It’s not eccentric, it’s not surprising. But it is a good electric car. A very interesting option if you are looking for a good family car as the only vehicle at home. And the Kia EV5 does not have the imprint and footprint of the EV9. Nor is it committed to that monolithic aspect of the EV3 that makes it so particular and that polarizes opinions about its design so much. This intermediate option seems like a kind of softened version of both cars without losing that muscular appearance, playing with straight and very pronounced edges. Its appearance, in fact, makes it appear larger. Its 4.61 meters seem to be more when you have it in front of you for the first time. We are, however, at figures very much to the taste of the European customer, who in this type of car largely opts for vehicles slightly larger than four and a half meters. With a wheelbase of 2.75 meters, the space for the rear seats is very good and maintains a trunk that, adding a front space in which little more than the charging cables can fit, reaches 566 liters. In the front area, it maintains the aesthetics and layout that has been accompanying the brand’s latest launches. The instrument panel and the central screen are embraced by the same frame, with a third digital space that unites both surfaces. All of this is supported on a kind of very clean horizontal desktop with touch buttons on the surface. On the steering wheel and the central area we have a multitude of physical buttons with some details that we liked. The instrument panel is displayed on a widely configurable 12.3-inch screen in its central area. In it we can find graphics of all kinds, from consumption to navigation or what the infotainment system is playing. Above the view we have a clear Head-Up Display with precise information for driving. The central screen, compatible with Android Auto and Apple CarPlayit is also 12.3 inches. Here, the possibilities are very wide and it has interesting solutions, such as a vertically sliding widget that supports the information displayed by the browser. However, I have two problems. The first is that it has so many shortcuts and so many functions to customize that it forces you to overcome a certain learning curve to be clear where each function is. I, who hadn’t gotten into a Kia for a while, had to spend some time finding, for example, the consumption data. My second problem is in the representation of the icons and shortcuts. The black background is useful to avoid confusing the driver but I think there is a lack of contrast in the icons. I, at least, have had some difficulty reading them clearly. I would have to test the car further to see if this can be fixed by, for example, increasing the screen brightness. Between both screens there is a third space in which the air conditioning is controlled. It seems like a good one to me. We have the basic … Read more

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.