DeepSeek

Anthropic just accused DeepSeek and other Chinese companies of “distilling” Claude

by

For months we have talked about the race between the United States and China to dominate artificial intelligence as if it were only a question of who trains the most powerful model or launches the next version first. But the pulse begins to move to another, more delicate area: that of the rules of the game. When one laboratory accuses another of extracting capabilities from its system to accelerate its own development, the discussion goes beyond the technical. That’s exactly what Anthropic just did by denounce “distillation” campaigns against his model Claude. The complaint. In a text published this Monday, the company claims to have detected “industrial-scale campaigns” aimed at extracting Claude’s capabilities. According to their version, the activities attributed to DeepSeekMoonshot and MiniMax reportedly involved more than 16 million queries, question and answer interactions, and were channeled through approximately 24,000 fraudulent accounts, in violation of their terms of service and regional access restrictions. The race and the suspicion. The announcement by the firm led by Darío Amodei occurs in a context of growing tension around the progress of Chinese AI. Let us remember that DeepSeek altered the Silicon Valley landscape a year ago with the launch of R1, a competitive model that was presented as Developed at a fraction of the cost of American alternatives. The impact was immediate on the markets and revived the political debate in Washington about the technological advantage over China. Distilling is not always cheating. Anthropic itself recognizes that distillation is a common technique in the sector. It consists, in simple terms, of training a less capable model using the responses generated by a more powerful one, something that large laboratories use to create smaller, cheaper versions of their own systems. The problem, according to the company, appears when this practice is used to “acquire powerful capabilities from other laboratories in a fraction of the time and at a fraction of the cost” that developing them independently would entail. In that case, distillation would cease to be an internal optimization and would become, always according to Anthropic, a way of taking advantage of the work of others. Recognizable pattern. The three laboratories would have used fraudulent accounts and proxy services to access Claude on a large scale while trying to avoid detection systems. The company details infrastructures, what it calls “hydra cluster”, extensive networks of accounts that distribute traffic between its API and third-party cloud platforms, so that when one account was blocked, another took its place. Anthropic maintains that what differentiated these activities from normal use was not an isolated query, but rather the massive and coordinated repetition of requests aimed at extracting very specific capabilities from the model. Three campaigns. Although Anthropic presents the campaigns as part of the same dynamic, it distinguishes relevant nuances. DeepSeek would have focused its more than 150,000 queries on extracting reasoning capabilities and generating safe alternatives to politically sensitive questions. Moonshot, with more than 3.4 million queries, would have been oriented towards the development of agents capable of using tools and manipulating computing environments. MiniMax would concentrate the largest volume, more than 13 million queries, and according to Anthropic’s account, it reacted in a matter of hours to the launch of a new system, redirecting its traffic to try to extract capabilities from its most recent system. A geopolitical issue. The company states that illicitly distilled models may lose safeguards that seek to prevent state or non-state actors from using AI for purposes such as the development of biological weapons or disinformation campaigns. It also argues that distillation undermines export controls by allowing foreign laboratories to close the gap in other ways, while at the same time recognizing that executing these large-scale extractions requires access to advanced chips, thus reinforcing the logic of restricting their availability while, at the same time, remembering that the risk would grow if these capabilities end up being integrated into military, intelligence or surveillance systems. Images | Xataka with Nano Banana Pro In Xataka | Seedance is the greatest brutality we have seen generating video. And it has an uncomfortable message: it has surpassed Sora and Veo without NVIDIA chips

DeepSeek is gaining users where the US has the most difficulty

by

about a year ago DeepSeek appeared on the radar of many people in the loudest way possible, with an impact that was noticed even on Wall Street. If the name sounds familiar to you, it comes from there. The interesting thing is that, twelve months later, its weight in the public conversation no longer seems the same, but that does not mean that it has disappeared from the board. In parallel, and according to the diagnosis that Microsoft now proposes, the Chinese startup continues to gain traction. The success of DeepSeek is worrying in the US. The warning comes from within the American ecosystem itself. Microsoft has warned that US AI groups face growing pressure from Chinese rivals in the battle for users in several markets, precisely because of the combination of “open” models and low prices. The winning strategy. What explains DeepSeek’s expansion has less to do with marketing and more to do with accessibility. The Redmond giant maintains in its report ‘Global AI Adoption in 2025‘that the company has reduced barriers to entry by offering a free chatbot on web and mobile, an especially attractive combination in cost-sensitive markets. DeepSeek also makes money. It is worth clarifying this so as not to be fooled: just because the chatbot is free does not mean that it does not have a business model. The firm founded by Liang Wenfeng distributes its technology with an open approach, with code under the MIT license and a separate licensing scheme for model weights. And, as is the case with most players in this industry, monetization is usually in the professional field: API accessthe interface that allows developers and companies to integrate these models into their own applications and services, is where much of the economic value is concentrated. Microsoft Map with Estimated DeepSeek Market Share The adoption map. The analysis itself places DeepSeek’s growth far from the markets where the technological narrative is traditionally decided, and breaks it down into two types of scenarios: emerging countries and countries where US services are limited or restricted. According to usage data, it is estimated that the Chinese group would have around 18% share in Ethiopia and 17% in Zimbabwe. And where American technological products are limited or restricted, the advance would be even greater, always according to these estimates: 56% in Belarus, 49% in Cuba and 43% in Russia. Target: Africa. Brad Smith, president of Microsoft, stated in an interview with the Financial Times thatif AI is to be deployed in Africa at scale, the problem is not just the software, but the infrastructure that supports it. According to their analysis, many African countries will need investment to build data centers and, in addition, mechanisms to subsidize the cost of electricity, one of the major operational limits. And here he introduces a relevant point: if the race depends solely on private capital, “it will not be enough” to compete with companies backed with a level of subsidy like the one that, he maintains, Chinese companies frequently have. A success that is still being measured. In essence, this case leaves a fairly clear idea: although DeepSeek sounds less popular today than it did a year ago, its approach is having a real impact in markets where it is not so easy for large American technology companies to deploy. It is an expansion that is driven more by accessibility than by narrative, and that is why it is also difficult to follow it from the West, until the data begins to appear. From here, the most interesting thing will be to see what happens in 2026: if DeepSeek manages to sustain that advantage and what other Chinese models, pushed by the same combination of openness, price and internal support, decide to follow in its wake. Images | Xataka with Gemini 3 Pro | Screenshot In Xataka | Anthropic has rewritten his 25,000-word “Constitution” for Claude. It is the manual for how AI should behave

it shoots up 500% and makes the creator of DeepSeek gold

by

Beijing’s quest for technological self-sufficiency has a new king: Moore Threads, the chip designer, has staged a historic stock market debut in Shanghai. Its shares soared more than 500% on its first day of trading. The euphoria has validated the strategy of a giant which, despite being on the US blacklist, has become one of the great hopes for breaking the semiconductor blockade. And in this maneuver, the great beneficiary has been the founder of DeepSeek. A debut and million-dollar profits. The IPO has not followed the usual channels. The China Securities Regulatory Commission gave the green light to the operation in just four months, a record time compared to the usual 470 days on average, something that underlines the state’s urgency to capitalize on the sector. According to SCMPLiang Wenfeng – through his fund – acquired more than 82,000 shares before the premiere. The result: a profit of almost $5.6 million in 48 hours. Nikkei Asia confirms that the company has reached a capitalization of 305 billion yuan (about $42 billion), becoming the fourth most valuable company on the STAR market. And it is not yet profitable: it hopes to be profitable in 2027. The pedigree of the alternative. The market is not buying just anything, it is buying the Chinese alternative to NVIDIA. Moore Threads is not just another startup; was founded in 2020 by Zhang Jianzhong, who was general manager of NVIDIA in China. In fact, this insider knowledge is what led the US to consider it a direct threat and include it on its blacklist in 2023. Its GPUs, such as the MTT S4000, are the spearhead of an industry that seeks to replace the H100 and H200—the latter yes it will arrive in China directly— Americans in state data centers, where the government already requires a 50% share of local chips for these crucial teams. It’s not just chips, it’s software. What makes Moore Threads dangerous to Jensen Huang’s business is not just the silicon, but its attack on an important technology for NVIDIA: CUDA. The Chinese startup has developed MUSA, a platform that allows you to recycle code written for NVIDIA and run it on your own GPUs. It is something that eliminates the main barrier to entry for Chinese companies that wanted to migrate but were trapped in the American software ecosystem. And it is also the missing piece in the puzzle of the historic alliance of Chinese companies forged to overthrow NVIDIA. The circle closes. The DeepSeek creator’s investment in Moore Threads is not reduced to financial terms. DeepSeek, which already hinted in August that I would no longer need NVIDIA chipsis collaborating closely with the chipmaker to optimize its AI models on domestic hardware. With an alternative to NVIDIA that triples its value and an AI capable of competing with Gemini and ChatGPT, China is building a closed ecosystem where hardware and software feed each other. It is a symbiosis that, in addition to uniting, shields. To the Chinese industry against any future sanctions from Washington. Cover image | Composition with images of Moore Threads and Matheus Bertelli for Pexels In Xataka | Cambricon Technologies: this company is China’s punch on the table to beat the US in AI

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

There is already a first crack in Chinese technological optimism: DeepSeek

by

Chen Deli, senior researcher at DeepSeek, has admitted at a state conference who is “extremely positive about technology, but pessimistic about its impact on society.” It is the first time that a representative of the Chinese company has spoken publicly since February, when its founder met with Xi Jinping after provoking that world earthquake with the launch of R1. And he has done it with that pessimistic outlook. Why is it important. This message comes from a company that the Chinese government has turned into a symbol of technological capacity and resilience in the face of US sanctions. That one of its leaders recognizes great risks for employment is a notable turn in a country where the official discourse is usually triumphalist. The facts. Chen participated in the World Internet Conference in Wuzhen along with the heads of five other companies known in China as “the six little dragons” of AI. His diagnosis has a gloomy tone: in one or two years, AI will be good enough to start replacing human jobs. In a decade or two it could take care of the rest. “Society could face an enormous challenge,” has said. “Tech companies need to take on the role of advocate.” Between the lines. This is not an American CEO peddling apocalypse smoke to inflate his valuation. In China, the State regulates technology with a firm hand. When Sam Altman says that AI will “probably lead to the end of the world, but in the meantime there will be big companies,” it sounds like marketing. When a DeepSeek executive says it at a conference organized by the government, after many months of silence and after its founder met with Xi, it sounds like a party line. The context. DeepSeek exploded in January with DeepSeek-R1a low-cost, open-source language model that was on par with American leaders. Since then, absolute exit. The founder, Liang Wenfeng, has appeared only once in all this time: at a televised symposium with Xi Jinping in February. Neither Liang nor the company has made public comments since then, and they have skipped all major Chinese tech conferences. Yes, but. While sending this message of caution, DeepSeek is in the process of consolidating itself as a cornerstone of the Chinese AI ecosystem. Chip manufacturers such as Cambricon and Huawei have developed hardware compatible with their models. In September, the company launched an “experimental” version of its V3 modelnotable not so much for its efficiency as for creating an alternative to NVIDIA’s CUDA API and its support for Chinese GPUs. In August, the simple announcement of a model optimized for national chips shares of the sector skyrocketed in the local market. And now what. Xi Jinping has proposed a little over a week ago on the APEC forum that there should be a global body that governs AI, making it “a public good for the international community.” Now a DeepSeek representative talks about AI as a potential threat that requires a unified approach from the technology sector. The narrative is shifting from triumphalism to preventive regulation. Featured image | Xataka, DeepSeek In Xataka | We believed that no open model could outperform GPT-5. A Chinese startup proves us wrong

they use Huawei and DeepSeek chips

by

China’s race to get become technologically independent from the United States It is reaching the military sector. The military is accelerating the integration of artificial intelligence into its operations and most importantly: they are favoring national technologies. In the software, DeepSeek. In hardware, Huawei chips. what’s happening. the chinese army is using AI to support strategic decision making and target detection. According to an analysis of Reutersseveral studies and patents suggest that they are also applying it in new vehicles such as robot dogs and autonomous drones, all prioritizing the use of national technologies, both in software and hardware. Why is it important. China has already given steps to stop depending on Nvidiathe maker of the most powerful AI chips. This is one more step towards technological independence, but in a critical sector such as the military. The objective is to eliminate foreign influence in its defense infrastructure, just like the United States does. Huawei chips. Speaking to Reuters, the defense policy expert Sunny Cheungassures that since the beginning of this year the Chinese military has increased the number of contractors that exclusively use national hardware. That is to say, AI chips made by Huawei. Although the military still uses Nvidia chips (it is not known if they were imported before or after of the blockade), there is a movement towards the use of own chips. DeepSeek. At the beginning of the year, military experts in China assured that the military was testing DeepSeek integration. In May, researchers from Xi’an University showed a system based on DeepSeek capable of creating and analyzing 10,000 combat scenarios in just 48 seconds. Reuters analyzed several tenders awarded to various companies by the Chinese military and at least a dozen mentioned DeepSeek, while only one referenced Alibaba’s Qwen. It is clear which is the preferred model for the Chinese army. Robot dogs and drones. The documents analyzed by Reuters also suggest that the Chinese military is integrating AI into autonomous vehicles such as robot dogs. It is no secret, in 2024 the army itself published a video promoting robot dogs who moved in packs to eliminate explosives and other threats. The robots in the video were from the Chinese company Unitree, but there are also other national companies dedicated to the manufacturing of these vehicles such as Norinco, which confirmed in a technical report that they use Huawei chips. On the other hand, Deepseek is also being integrated into drones to give them the ability to recognize and follow targets with hardly any human intervention. Image | Wikipedia, Flickr In Xataka | Europe already has the future of war drones within its reach. And it is offered by a country accustomed to them: Israel

Hangzhou is the city of DeepSeek, Alibaba and Unitree without any of the typical Silicon Valley ingredients. His secret is another

by

Hangzhou, a city of 12 million inhabitants 180 km south of Shanghai, is home to a striking number of powerful technology companies: Seven reference technologies (the six ‘little dragons’ plus the giant Alibaba) in a city that does not have any of the elements considered essential in Silicon Valley: Abundant venture capital. Leading universities. Links between university and industry. And a robust industrial structure. How could you then Hangzhou emerge well? The facts. Venture capital is plummeting in China. Funds in yuan have fallen from 88.42 billion dollars in 2022 to 5.38 billion in 2024. Funds in dollars, from 17.32 billion to 750 million. Hangzhou has not been a major recipient of investment until last year, when its province –Zheijang– stood out with 41 new corporate venture capital funds. But it was only after Unitree or Game Science had gained national attention. Missing. Hangzhou has only one elite university – Zhejiang – compared to 26 in Beijing, 11 in Jiangsu or 10 in Shanghai. The admission rate at Tsinghua and Beijing Universities for students from the capital (0.85%) is almost ten times that of students from Zhejiang (0.09%). None of the founders of “the six little dragons” or Alibaba created their company directly from university. Liang Wenfeng founded High-Flyer, the hedge fund after DeepSeekeight years after graduating. Jack Ma was rejected for 30 jobs after finishing his studies. Yes, but. The city has innovated by doing away with those ingredients. The explanation offered by Zilan Qian, a researcher at the Oxford China Policy Lab, points out ChinaTalk to “flexible governance”: a model where officials adopt “waiters” and “babysitters” mentality that facilitate rather than control. The context. Hangzhou does not have the political, financial or industrial importance of first-tier cities, which has given it greater local autonomy to shape its technology sector. Zhejiang province was a pioneer since the 1980s in promoting private enterprise during the early phases of Chinese economic reforms. Jack Ma He tried to establish Alibaba’s headquarters in Beijing or Shanghai, but failed due to the cost of rent and bureaucratic barriers. In 2015, Ma explained her decision: “Beijing favors state-owned enterprises, Shanghai favors foreign companies, and Alibaba was nothing in their eyes. If we return to Hangzhou, we become the local only child who receives all the attention and support.” Hangzhou is part of the sometimes called “chinese technology triangle“(sometimes also”golden triangle“) along with Shanghai and Shenzhen. More than a geometric reality, the functional metaphor describes the complementarity of three cities: Shenzhen provides industrial capacity and hardware. Shanghai concentrates finances and internationalization Hangzhou stands out in the internet, AI and an ecosystem favorable to private companies. Each vertex of the triangle has different strengths that, combined, generate an ecosystem where geographical proximity facilitates collaboration and flow of talent between the three poles. Between the lines. The model is described as “market-oriented” but maintains a level of centralized governance. The Hangzhou government sees quality of life as a strategy to attract businesses and talent, but positions itself as an enabler, not a controller. The absence of state-backed research institutes and a strong industrial base contributes to the government’s humble attitude. If Hangzhou were more strategic or more industrial, DeepSeek might not have had the creative space to emerge and provoke the earthquake that caused in January. The narrative of “self-made industry” and “entrepreneurial bureaucracy” admits conflicting readings. What some interpret as facilitation, others read as a euphemism for “dirigiste intervention by the State”, with a very defined plan of action and long-term objectives. “Flexible governance” can be both real local autonomy… and dirigisme disguised as pragmatism. At least it is no longer “a city south of Shanghai” but “Alibaba City” or “DeepSeek City”. In Xataka | China is selling us a future full of humanoid robots. We have (many) doubts Featured image | JinHui CHEN in Unsplash

Deepseek put China on the AI ​​map. The danger is that this revolution stays in a day flower

by

Deepseek R1 was eating the world At the beginning of the year. This Chinese model, apparently out of nowhere, caused A true shock In the AI ​​industry, but since then there has been movement. Actually there has been one, but the disturbing thing is precisely what that movement has been. Hi, Deepseek v3.1. The startup advertisement Last week the launch of Deepseek V3.1, a new version that stood out for being an improved hybrid of Deepseek V3 (fast response) and Deepseek R1 (reasoning). There was also good news in terms of their performance: according to the Benchmarks published by those responsible, it was significantly higher than their predecessors. Visible (but non -dramatic) improvements. In the “model card” (model card) that those responsible offers In Hugging FaceDeepseek v3.1 (in reasoning mode) proved to behave slightly better than Deepseek R1-0528, —Your previous version, more powerful-in areas such as programming or in mathematical tests, but some users who have tried it there comment That except in those areas, the model is worse and “it behaves poorly when following instructions or prompts provided by users.” Others confirm it and They assure which is useful for programmers, but not for other areas. It also has limitations on its multimodal support, and focuses on the text instead of providing more options for another type of interaction, for example from voice, image, video or audio messages. A Chinese model for Chinese chips. But even more interesting it was that Deepseek V3.1 has been designed and launched with a clear objective: avoiding the dependence of foreign chips. The FP8 precision used makes this model behave very well In the next -generation Chinese chips. The strategy seems very interesting for the startup, which could thus have a very aligned model with the priorities of the Chinese government. This is: use local models for local chips as much as possible. And R1, what? From there some doubts arise. The first, which affects Deepseek R1, the model with which the startup “broke” the market at the beginning of the year. The company has eliminated all references to this model in the characteristic of “deep thought”, which has generated doubts about the potential appearance of its expected successor, a hypothetical Deepseek R2. Loses users. But while that theoretical model comes – if it does – the company faces a more immediate threat. As they point out In SCMPDeepseek is losing users (or at least relevance) in recent months. In the first quarter of the year its market share within the scope of the IA Open Source models used on the PPIO cloud platform was a spectacular 99%. However, in the second quarter that percentage has dropped to 80%. Fierce competition. That fall relevance has an obvious reason: its local competitors are squeezing. And a lot. Among them is the family of models Qwen from Alibaba, but Also others like Kimi-K2-Instructof the startup Mosohot AI – in which Alibaba has also invested – which is becoming one of the most popular models of recent weeks. Delays and deceleration. Precisely the focus on being able to make the most of future Chinese chips seems to be the reason that this hypothetical Deepseek R2 is being delayed. At least that is the hypothesis that consider In Financial Timeswhere they revealed that the startup has failed when trying to train with Huawei chips. The situation has made them Training with Nvidia chipsand that are using the Huawei Asce for the inference stage, that is, the interaction with the model via web or API by users. But this attitude is “very Chinese”. We may in Western countries we are accustomed to a much more frantic pace and that we expect constant updates and improvements with an eye on the short term. In China, philosophy is usually the opposite, and companies adopt A long -term strategy even if immediate benefits are lost. Maintaining a low profile is also usual among those companies, which try not to make much noise … until they do, as Depseek has already demonstrated. Thus, we will have to remain very attentive to the activity of this startup, because surely he will be working to continue being one of the protagonists of the AI ​​panorama. Image | Tim Reckmann In Xataka | Deepseek has suggested that Nvidia chips no longer needs. We believe to know who is buying them

Deepseek has suggested that Nvidia chips no longer needs. We believe to know who is buying them

by

Deepseek shook at the beginning of 2025 the foundations of the industry of the artificial intelligence (AI). This Chinese model developed by the specialized quantitative coverage fund trading High-Flyer algorithmic monopolized all the attention because He gave us free of charge an AI with a quality similar to the comparable solutions of Openai or Google. However, this was not all. And it is that for three weeks the debate about the hardware used by this company to train its model. High-Flyer, who specializes in addressing investment decisions using advanced mathematical models and computational algorithms, says he trained Deepseek R1 using 2,048 chips H800 of Nvidia. Several analysts They soon react ensuring that he had actually used 50,000 GPU H100which are more powerful, bought through intermediaries. The sanctions of the US government prevent Nvidia According to SCMPDeepseek’s next iteration will dispense with the GPUs of this American company. High-Flyer has suggested that everything you need already finds it in China Last week those responsible for the Deepseek development published on the Wechat platform an entry in which they hinted that next generation GPUs for China They will launch soon. This comment has unleashed endless speculation about the name of the Chinese company that is going to make this announcement, but, above all, it has put on the table the possibility that High-Flyer is preparing the following Deepseek iteration using these chips. A priori does not seem at all crazy. According to SCMP The five GPU designers for the Chinese who have the ability to deliver to High-Flyer the hardware they need are Huawei Technologies, Cambricon Technologies, Moore Threads, Hygon Information Technology and Metax Integrated Circi. Any of them could be responsible for the announcement anticipated by Deepseek developers in Wechat, but we We bet on the first three Because, in our opinion, they are Those who are “in a better way”. Cambricon Technologies is one of the companies specialized in the design of GPU for AI with greater growth potential Although it is not as well known as Huawei or Moore Threads, Cambricon Technologies is one of the companies specialized in the design of GPU for AI with greater growth potential. In fact, he has received the approval of the Shanghai bag (China) to raise 560 million dollars. Will allocate them to the design of four chips for training and inference of AI models, and also to the development of an alternative to CUDAfrom Nvidia. On the other hand, Moore Threads He has developed several GPU for AI applications that, on paper, rivaize some of the advanced solutions that have placed in the Nvidia, AMD or Huawei market. MTT S4000 and MTT S3000 cards They are its most interesting proposals right now, although, curiously, in its porpholio the MTT S80 card also appears, a proposal for games and content creation that, according to Moore Threads itself, has a calculation capacity of 14.4 Tflops in single -precision floating coma operations. The other indispensable actor in the Chinese chips industry for IA is Huawei. His most ambitious proposal right now is the chip Ascend 910dwho seeks to overcome the performance of the GPU NVIDIA H100. However, this Chinese company has also recently presented its chip Ascend 920a solution that is clearly destined to occupy in the Chinese market the gaps that the NVIDIA H20 GPU is going to leave. This proposal will enter large -scale production during the second half of 2025 using 6 NM integration technology that have presumably developed elbow with Huawei elbow and SMIC (Semiconductor manufacturing international corp). More information | SCMP In Xataka | Nvidia has to deal with the absolute distrust of several US legislators. His plan in China is in danger In Xataka | The US wants to end the chips for the Chinese that are sold abroad. And China knows how to defend oneself

Deepseek marked a turning point in the AI race. Now another Chinese company wants to imitate its success: Kimi K2 is born

by

The Chinese startup Monshot AI has presented Kimi K2, an open -source artificial intelligence model that arrives with outstanding programming capabilities and autonomous tasks that, according to The published benchmarksThey spray competition in several of their models. Its launch occurs at a key moment for the sector, when Chinese companies seek to replicate the disruptive success of Deepseek with potential height models and much cheaper than market alternatives. Kimi does not come from nothing. MoNshot ai was one of the most promising startups in the Chinese ecosystem of AI and that giants like Alibaba have invested greatly. His Kimi chatbot reached third place in monthly active users in August 2024, but fell to the seventh in June After the emergence of Deepseek R1 in January. Now try to recover ground with a strategy that combines open source and aggressive prices, following the formula that catapulted Deepseek. Image: MoNshot AI What Kimi K2 offers. The model has 1 billion total parameters and 32,000 million activated parameters, using The well-known Mixture-Of-Experts architecture to optimize computational costs. It is presented in two versions: a base for researchers and developers, and another optimized for conversation and autonomous tasks. Kimi K2 thus becomes Moonshot AI’s proposal with the ability to act as an intelligent agent to use tools, write code, complete workflows or talk, among other tasks. Kimi K2 explained in numbers. In performance testsKimi K2 has achieved 65.8% precision at Swe-Bench Verified, one of the most demanding benchmarks for software engineering. In LivecodeBench it reached 53.7%, exceeding 46.9% of Deepseek-V3 and 44.7% of GPT-4.1. In mathematics, its 97.4% score in Math-500 exceeds 92.4% of GPT-4.1, suggesting significant advances in mathematical reasoning. The price factor. MoNshot is charging $ 0.15 per million input tokens and $ 2.50 per million tokens out of the developers who use their API. Compared, Claude Opus 4 It charges 100 times more for the entrance (15 dollars) and 30 times more for the output ($ 75), while GPT-4.1 charges 2 dollars per entrance and 8 per exit. In addition, the model is available for free in Web applications and Kimi mobile, without monthly subscriptions that require chatgpt or Claude for their most advanced models. Technical innovation. MoNshot has developed the MuCanclip optimizer, which allows train models of one billion parameters “With zero training instability.” This technology could drastically reduce the training costs of large models, a problem that has limited the development of AI to companies with greater resources. Double channel strategy. The company offers so much Free access to the source code as payment API at a very competitive price. This strategy allows companies to start with the API for immediate implementation and then migrate to self -healing versions either by regulatory cost or compliance. And it is that each developer who downloads Kimi K2 becomes a potential business client. Moment of inflection. Kimi K2 represents a convergence point where open source models and proprietary alternatives shake hands. MoNshot AI intends to turn Kimi into a tool for everything, while offering its open source model and is reserved to charge for the use of its API for all types of implementations. And now what. The launch reaches a critical point in which both Openai, such as Google or Anthropic, must respond to this wave of cheap and high quality language models. The issue is no longer whether open source models can match the owners, but if large technological ones can adapt their business models fast enough to compete in this new scenario. The looks are put in GPT-5 And in the next movements of the industry at a rate, as always, accelerated. Cover image | Xataka with Mockuuuups Studio and Kimi AI In Xataka | Grok 4 destroys the tests and aims to be the most advanced AI model. The problem is that Elon Musk continues to sabotage his answers

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.