In the last week of January 2025, a small artificial intelligence laboratory in the eastern Chinese city of Hangzhou released, in quick succession, two open-source large language models that, within a matter of days, would upend the global conversation about the cost, the ownership, and the future of frontier AI. The laboratory was DeepSeek, a year-old company that had been spun out of a Chinese quantitative hedge fund called High-Flyer, and the two models were DeepSeek R1, a reasoning model that could match OpenAI’s o1 on a wide range of mathematical and scientific benchmarks at a tiny fraction of the training cost, and DeepSeek V3, the base model on which R1 had been built, a 671 billion parameter mixture of experts model that had been trained for an order of magnitude less than the equivalent American frontier models. The release, in the late hours of a winter Monday in late January 2025, was not, in the immediate aftermath, a major media event in the United States or in Europe: DeepSeek was a small company, the release was on a Friday night in China, and the immediate technical response, while positive, did not suggest that anything dramatic had happened. The dramatic thing happened the following Monday, January 27 2025, when the American stock market opened and the share price of Nvidia, the American chipmaker whose graphics processing units, GPUs, were the principal hardware used in training frontier AI models, fell by nearly seventeen percent in a single trading session, losing approximately six hundred billion dollars of market capitalisation, the largest single-day drop in the history of the American stock market. The reason for the drop, in the immediate market interpretation, was DeepSeek: if a small Chinese lab could train a frontier model on a tiny fraction of the cost of the American frontier labs, then the demand for the American AI infrastructure, and in particular for the high-end Nvidia GPUs, was going to be much smaller than the market had assumed. The DeepSeek moment, in the financial and strategic interpretation that took hold in the weeks that followed, was a Sputnik moment for American AI, a moment in which the assumption of American technological dominance in artificial intelligence was, for the first time, openly questioned. The DeepSeek moment was also, in the technical interpretation, a moment in which the open-source Chinese AI ecosystem, which had been growing quietly in the years since 2023, was, for the first time, openly recognised as a major force in the global AI ecosystem. The DeepSeek moment was, in sum, a turning point in the history of artificial intelligence, and a turning point in the history of the relationship between the United States and China in the global AI race. To understand the DeepSeek moment, it is necessary to understand the company that produced it, the technical choices that made the moment possible, the strategic choices that made the moment consequential, and the implications that the moment has had for the global AI ecosystem in the months and years since.
From High-Flyer to DeepSeek: the hedge fund origins of a frontier lab
DeepSeek is, in many respects, an unusual company. The company was founded in 2023 by Liang Wenfeng, a soft-spoken thirty-eight year old Chinese entrepreneur who had made a substantial fortune as the founder of High-Flyer, one of the largest and most successful quantitative hedge funds in China. Liang was born in 1985 in Zhanjiang, a port city in the southern Chinese province of Guangdong, and he was educated at Tsinghua University, the principal Chinese institution for the study of computer science and engineering. At Tsinghua, Liang studied information engineering and then computer science, and he was, by the standards of his Tsinghua cohort, a notably talented student, with a particular interest in machine learning and in the application of machine learning to the analysis of large financial datasets. After graduation from Tsinghua, Liang worked briefly as a research engineer at a small Chinese AI laboratory, and then, in 2013, he founded a quantitative trading firm, initially called Magic Cube, that was, in the years that followed, renamed High-Flyer. High-Flyer was, from the beginning, an unusual quantitative trading firm. The firm was, in the first instance, one of the first Chinese quantitative trading firms to make a serious investment in machine learning: while most of the Chinese quantitative trading firms of the period relied on more traditional statistical methods for the analysis of financial data, High-Flyer was, from the beginning, a machine-learning-first firm, with a substantial research team devoted to the development of new machine learning methods for the analysis of financial data. The firm was, in the second instance, one of the first Chinese quantitative trading firms to make a serious investment in computational infrastructure: High-Flyer built, in the years between 2016 and 2023, a substantial internal computing cluster, initially for the purpose of training machine learning models on financial data, and the cluster was, by 2021, one of the largest commercial computing clusters in China, with a substantial fraction of the cluster devoted to the development of new machine learning models for the analysis of financial data.
The pivot to general-purpose AI, in the High-Flyer story, came in 2021, when Liang Wenfeng decided to invest a substantial fraction of the High-Flyer capital in a new project, internally called Fire-Flyer, that was, in the original conception, a project for the development of a large language model for use in the High-Flyer trading operation. The Fire-Flyer project was, in the first instance, a research project: a small team of High-Flyer engineers and machine learning researchers, working in collaboration with a small number of Tsinghua University computer science faculty, was charged with the development of a large language model that could, in the long run, be used in the High-Flyer trading operation. The Fire-Flyer project was, in the second instance, an infrastructure project: the project was, in the years between 2021 and 2023, the principal justification for a substantial expansion of the High-Flyer computing cluster, and the expansion, in turn, gave the project a substantial computational resource that was, by the standards of the period, very large. The Fire-Flyer project was, in the third instance, a recruitment project: the project was, in the years between 2021 and 2023, the principal vehicle for the recruitment of a substantial number of top Chinese machine learning researchers, many of them recent PhD graduates from Tsinghua and from Peking University, and the recruitment gave the project a substantial intellectual resource that was, by the standards of the period, also very large.
The High-Flyer story, in the broader context of the Chinese quantitative trading industry, was, in the years between 2013 and 2023, a particularly instructive example of the rise of the Chinese quantitative trading industry. The Chinese quantitative trading industry, in the years between 2013 and 2023, had grown from a small niche within the broader Chinese asset management industry to a substantial fraction of the Chinese asset management industry, with the largest Chinese quantitative trading firms, including High-Flyer, the Ningbo-based firm Jimuquant, the Shenzhen-based firm Ubiquant, and the Shanghai-based firm Magic Bridge, all of which had, in the years between 2013 and 2023, grown to be substantial firms with multi-billion dollar assets under management. The Chinese quantitative trading industry, in the years between 2013 and 2023, was, in addition, a particularly innovative industry, with the largest Chinese quantitative trading firms, including High-Flyer, having been early adopters of the new machine learning methods that were, in the years between 2013 and 2023, being developed in the broader machine learning research community. The Chinese quantitative trading industry, in the years between 2013 and 2023, was, in particular, a particularly important source of demand for high-performance computing hardware, with the largest Chinese quantitative trading firms, including High-Flyer, having been early customers of the high-end Nvidia GPUs that were, in the years between 2013 and 2023, the principal hardware used for the training of machine learning models. The High-Flyer story, in the years between 2013 and 2023, was, in sum, a particularly instructive example of the rise of the Chinese quantitative trading industry, and the High-Flyer story, in the years between 2013 and 2023, was a particularly important source of the technical and the financial resources that were, in the years that followed, going to be deployed in the DeepSeek project.
The High-Flyer computing cluster, in the years between 2016 and 2023, was, in the first instance, one of the largest commercial computing clusters in China, and the High-Flyer computing cluster was, in the years between 2016 and 2023, a particularly important source of the technical and the financial resources that were, in the years that followed, going to be deployed in the DeepSeek project. The cluster, in the first instance, was based on a substantial number of the high-end Nvidia GPUs that were, in the years between 2016 and 2023, the principal hardware used for the training of machine learning models, with the cluster having, in the years between 2016 and 2023, grown to include a substantial fraction of the high-end Nvidia GPUs that were, in the period, available to Chinese commercial customers. The cluster, in the second instance, was a particularly well-engineered computing cluster, with the cluster having, in the years between 2016 and 2023, been a particular focus of the High-Flyer engineering team, and the cluster, in the years that followed, being a particularly important source of the technical expertise that was, in the years that followed, going to be deployed in the DeepSeek project. The cluster, in the third instance, was a particularly well-managed computing cluster, with the cluster having, in the years between 2016 and 2023, been a particular focus of the High-Flyer operations team, and the cluster, in the years that followed, being a particularly important source of the operational expertise that was, in the years that followed, going to be deployed in the DeepSeek project.
The formal founding of DeepSeek came in 2023, when Liang Wenfeng decided to spin off the Fire-Flyer project as a separate company, with the company’s name being a combination of the words deep, in the sense of deep learning, and seek, in the sense of the search for new knowledge. The new company, DeepSeek, was based in Hangzhou, the capital of the eastern Chinese province of Zhejiang and the principal Chinese city for the development of the Chinese internet industry. Hangzhou was, in the years that followed, the home of Alibaba, the principal Chinese e-commerce and cloud computing company, and the home of a substantial ecosystem of Chinese internet companies, including NetEase, the principal Chinese gaming company, and a substantial number of smaller companies. The new company, DeepSeek, was, in the years between 2023 and 2024, a small but very well-funded company, with a substantial war chest inherited from the High-Flyer hedge fund and with a substantial computing cluster inherited from the Fire-Flyer project. The new company was, in addition, a small but very well-staffed company, with a substantial research team inherited from the Fire-Flyer project and with a substantial recruitment pipeline into the top Chinese computer science departments. The new company, in sum, was, in the years between 2023 and 2024, one of the best-resourced small AI laboratories in the world, and a laboratory that was, in the years that followed, one of the most prolific open-source AI laboratories in the world.
The early DeepSeek story, in the years between 2023 and the end of 2024, was, in the immediate aftermath, a story of a small but rapidly maturing laboratory. The laboratory, in the period, released a series of progressively more capable open-source large language models, including the DeepSeek LLM 67B base model in November 2023, the DeepSeek-V2 mixture of experts model in May 2024, the DeepSeek-Coder V2 code model in June 2024, the DeepSeek-V2.5 model in September 2024, and the DeepSeek-V3 base model in December 2024. The laboratory, in the period, was, in addition, a small but very well-regarded laboratory in the global open-source AI community, with a substantial following on Hugging Face, the principal platform for the distribution of open-source AI models, and a substantial following on GitHub, the principal platform for the distribution of open-source software. The laboratory was, in the period, recognised as a major force in the open-source AI ecosystem, but the laboratory was, in the period, not yet a major force in the broader AI conversation, which was still, in the years between 2023 and the end of 2024, dominated by the closed-source American AI laboratories, including OpenAI, Anthropic, and Google. The DeepSeek moment, when it came in January 2025, was, in the immediate aftermath, a moment in which the laboratory, which had been quietly maturing for two years, was, for the first time, recognised as a major force in the broader AI conversation.
DeepSeek V3: the technical breakthrough
The DeepSeek V3 model, released in late December 2024, was, in the technical interpretation, a major step forward for the open-source AI ecosystem. The model was, in the first instance, a very large model: 671 billion parameters, of which approximately 37 billion were active for any given token, in a mixture of experts architecture in which the model was divided into 256 routed experts and a small number of shared experts, with each token being routed to a small subset of the routed experts. The model was, in the second instance, a very efficiently trained model: the model was trained on a corpus of approximately 14.8 trillion tokens, and the training was completed in approximately 2.788 million H800 GPU hours, which, at the then-current commercial rate of approximately two dollars per H800 GPU hour, would have cost approximately 5.576 million dollars. The training cost claim was, in the immediate aftermath, a major point of controversy: critics noted that the cost calculation excluded the substantial cost of the prior research, the cost of the data curation, the cost of the human evaluation, and the cost of the failed training runs that were, in any frontier training effort, a substantial fraction of the total cost. The cost claim was, however, in the rough order of magnitude of the training cost, and the cost was, by the standards of the American frontier AI laboratories, which were spending hundreds of millions to low billions of dollars on the training of each frontier model, a tiny fraction of the cost. The cost claim, in the technical interpretation, was a major part of the DeepSeek moment: if a frontier model could be trained for approximately 5.5 million dollars, then the assumption that the frontier of AI was reserved for the well-capitalised American AI laboratories was, in the immediate aftermath, openly questioned.
The technical architecture of DeepSeek V3 was, in the technical interpretation, a careful combination of established and novel techniques. The model used, in the first instance, the Multi-Head Latent Attention, MLA, mechanism that had been introduced in DeepSeek V2, a mechanism that was, in the technical interpretation, a major improvement over the standard Multi-Head Attention mechanism. The standard Multi-Head Attention mechanism, which was introduced in the original Transformer paper, requires the model to store the full key-value cache for each token, a cache that grows linearly with the sequence length and that, for long sequences, becomes a substantial fraction of the total memory and compute of the model. The Multi-Head Latent Attention mechanism, in contrast, compresses the key-value cache into a small latent representation, and the compressed representation is, in the technical interpretation, a major source of the efficiency of the model. The MLA mechanism was, in the technical interpretation, one of the principal technical contributions of the DeepSeek V2 model, and the MLA mechanism was, in the technical interpretation, a major reason for the efficiency of the DeepSeek V3 model. The model used, in the second instance, the DeepSeek MoE mixture of experts architecture, an architecture that was, in the technical interpretation, a major improvement over the standard mixture of experts architecture. The standard mixture of experts architecture, which was introduced in the early 1990s, divides the model into a number of experts, with each token being routed to a small subset of the experts, and the standard mixture of experts architecture is, in the technical interpretation, a major source of the efficiency of the model. The DeepSeek MoE architecture, in contrast, divides the experts into a number of routed experts and a small number of shared experts, and the shared experts, in the technical interpretation, are a major source of the stability and the quality of the model. The DeepSeek MoE architecture was, in the technical interpretation, one of the principal technical contributions of the DeepSeek V2 model, and the DeepSeek MoE architecture was, in the technical interpretation, a major reason for the quality of the DeepSeek V3 model.
The training of DeepSeek V3 was, in the technical interpretation, a careful combination of established and novel techniques. The training used, in the first instance, FP8 mixed-precision training, a training technique that was, in the technical interpretation, a major improvement over the standard FP16 or BF16 mixed-precision training. The FP8 mixed-precision training reduces the precision of the numerical representations used in the training, and the reduction, in the technical interpretation, is a major source of the efficiency of the training. The FP8 mixed-precision training was, in the technical interpretation, one of the principal technical contributions of the DeepSeek V3 training, and the FP8 mixed-precision training was, in the technical interpretation, a major reason for the efficiency of the training. The training used, in the second instance, the DualPipe algorithm, a novel pipeline parallelism algorithm that was, in the technical interpretation, a major improvement over the standard pipeline parallelism algorithms. The DualPipe algorithm reduces the communication overhead of the pipeline parallelism, and the reduction, in the technical interpretation, is a major source of the efficiency of the training. The DualPipe algorithm was, in the technical interpretation, one of the principal technical contributions of the DeepSeek V3 training, and the DualPipe algorithm was, in the technical interpretation, a major reason for the efficiency of the training. The training used, in the third instance, the Multi-Token Prediction objective, a training objective that was, in the technical interpretation, a major improvement over the standard next-token prediction objective. The Multi-Token Prediction objective requires the model to predict multiple tokens at each position, and the requirement, in the technical interpretation, is a major source of the quality of the model. The Multi-Token Prediction objective was, in the technical interpretation, one of the principal technical contributions of the DeepSeek V3 training, and the Multi-Token Prediction objective was, in the technical interpretation, a major reason for the quality of the model.
The training dataset of DeepSeek V3 was, in the technical interpretation, a particularly important source of the quality of the model. The training dataset was, in the first instance, a particularly large dataset, with the dataset having, in the technical interpretation, included approximately 14.8 trillion tokens of training data, and the dataset having, in the technical interpretation, been a particularly large dataset by the standards of the period. The training dataset was, in the second instance, a particularly diverse dataset, with the dataset having, in the technical interpretation, included a wide range of different types of training data, including web pages, books, academic papers, code, and conversational data, and the diversity of the dataset having, in the technical interpretation, been a particularly important source of the quality of the model. The training dataset was, in the third instance, a particularly well-curated dataset, with the dataset having, in the technical interpretation, been subjected to a substantial amount of data curation, including deduplication, filtering, and quality scoring, and the curation of the dataset having, in the technical interpretation, been a particularly important source of the quality of the model. The training dataset, in sum, was a major part of the DeepSeek V3 training, and the training dataset, in sum, was, in the technical interpretation, one of the principal technical contributions of the DeepSeek V3 training.
The evaluation of DeepSeek V3, in the technical interpretation, was a particularly extensive and a particularly rigorous evaluation. The evaluation, in the first instance, was based on a particularly wide range of standard benchmarks, including the MMLU benchmark, the MMLU-Pro benchmark, the MATH-500 benchmark, the HumanEval benchmark, the GSM8K benchmark, the BBH benchmark, the DROP benchmark, the HellaSwag benchmark, the PIQA benchmark, the SIQA benchmark, the WinoGrande benchmark, the ARC-Challenge benchmark, the TriviaQA benchmark, the NaturalQuestions benchmark, and a substantial number of other standard benchmarks. The evaluation, in the second instance, was based on a particularly wide range of human evaluations, including human evaluations of the model’s performance on a wide range of different tasks, including creative writing, technical writing, summarization, question answering, and code generation. The evaluation, in the third instance, was based on a particularly wide range of safety evaluations, including safety evaluations of the model’s performance on a wide range of different safety benchmarks, including the BBQ benchmark, the Toxigen benchmark, and the HarmBench benchmark. The evaluation, in sum, was a major part of the DeepSeek V3 release, and the evaluation, in sum, was, in the technical interpretation, one of the principal contributions of the DeepSeek V3 release to the open source AI community.
The performance of DeepSeek V3, on the standard benchmarks of the period, was, in the technical interpretation, broadly comparable to the closed-source American frontier models. The model, on the MMLU benchmark, a multiple choice benchmark covering a wide range of academic subjects, scored approximately 88.5 percent, a score that was, by the standards of the period, broadly comparable to the closed-source American frontier models. The model, on the MATH-500 benchmark, a benchmark covering a wide range of mathematical problems, scored approximately 90.2 percent, a score that was, by the standards of the period, broadly comparable to the closed-source American frontier models. The model, on the HumanEval benchmark, a benchmark covering a wide range of code generation problems, scored approximately 82.6 percent, a score that was, by the standards of the period, broadly comparable to the closed-source American frontier models. The model, on the GSM8K benchmark, a benchmark covering a wide range of grade-school mathematical word problems, scored approximately 89.3 percent, a score that was, by the standards of the period, broadly comparable to the closed-source American frontier models. The model, on the MMLU-Pro benchmark, a more challenging variant of the MMLU benchmark, scored approximately 75.9 percent, a score that was, by the standards of the period, broadly comparable to the closed-source American frontier models. The model, in the technical interpretation, was a major step forward for the open-source AI ecosystem, and the model was, in the technical interpretation, one of the principal technical contributions of the DeepSeek laboratory to the global AI ecosystem.
DeepSeek R1 and the January 2025 moment
The DeepSeek R1 model, released in late January 2025, was, in the technical interpretation, a major step forward for the open-source AI ecosystem. The model was, in the first instance, a reasoning model, a model that was specifically designed to be able to perform complex multi-step reasoning tasks, and the model was, in the technical interpretation, a major step forward over the standard chat models that had been the principal focus of the open-source AI ecosystem in the years before. The reasoning model was, in the second instance, a model that was built on top of the DeepSeek V3 base model, a base model that was, in the technical interpretation, one of the most capable open-source base models in the world, and the use of the DeepSeek V3 base model was, in the technical interpretation, a major reason for the quality of the DeepSeek R1 model. The reasoning model was, in the third instance, a model that was trained using a novel training methodology, a methodology that combined reinforcement learning with a small amount of high-quality supervised fine-tuning data, and the methodology was, in the technical interpretation, a major reason for the quality of the model. The methodology was, in the fourth instance, a methodology that was, in the technical interpretation, a major step forward over the standard reasoning model training methodologies, methodologies that had, in the years before, been the principal focus of the closed-source American AI laboratories.
The release of DeepSeek R1 was, in the technical interpretation, a major event in the open-source AI ecosystem. The model was released, in late January 2025, in two principal variants: the DeepSeek R1 model, the full reasoning model with all of the reinforcement learning training, and the DeepSeek R1 Zero model, a variant of the model that had been trained using pure reinforcement learning without any supervised fine-tuning. The DeepSeek R1 Zero model was, in the technical interpretation, a particularly notable release, because the model was, in the technical interpretation, the first major open-source reasoning model that had been trained using pure reinforcement learning, and the model was, in the technical interpretation, a major demonstration of the power of the reinforcement learning approach to reasoning. The DeepSeek R1 model was, in addition, a model that was, in the technical interpretation, broadly comparable in quality to the OpenAI o1 model, the closed-source American reasoning model that had been released in late 2024 and that was, in the technical interpretation, the principal reference point for the reasoning model benchmark. The DeepSeek R1 model, on the AIME benchmark, a benchmark covering a wide range of high-school-level mathematical olympiad problems, scored approximately 79.8 percent, a score that was, by the standards of the period, broadly comparable to the OpenAI o1 model. The DeepSeek R1 model, on the MATH-500 benchmark, scored approximately 97.3 percent, a score that was, by the standards of the period, broadly comparable to the OpenAI o1 model. The DeepSeek R1 model, on the MMLU benchmark, scored approximately 90.8 percent, a score that was, by the standards of the period, broadly comparable to the OpenAI o1 model.
The stock market reaction, in the days after the release of DeepSeek R1, was, in the financial interpretation, a major event in the history of the American stock market. The Nvidia share price, in the late hours of January 27 2025, fell by approximately seventeen percent, losing approximately six hundred billion dollars of market capitalisation, the largest single-day drop in the history of the American stock market. The reason for the drop, in the immediate market interpretation, was DeepSeek: if a small Chinese lab could train a frontier reasoning model on a tiny fraction of the cost of the American frontier labs, then the demand for the American AI infrastructure, and in particular for the high-end Nvidia GPUs, was going to be much smaller than the market had assumed. The drop was, in the immediate financial interpretation, a major repricing of the American AI industry, and a major repricing of the assumption that the American AI industry was, in the long run, going to be the dominant force in the global AI ecosystem. The drop was, in addition, a major shock to the American AI policy establishment, and a major shock to the American AI strategy establishment, both of which had, in the years before, assumed that the American AI industry was, in the long run, going to be the dominant force in the global AI ecosystem.
The specific reasoning capabilities of DeepSeek R1, in the technical interpretation, were a particularly significant development in the open source AI ecosystem. The model, in the first instance, demonstrated a particularly strong ability to perform complex multi-step mathematical reasoning, with the model having, in the technical interpretation, been able to solve a wide range of complex mathematical problems that were, by the standards of the period, beyond the capabilities of the standard open source AI models. The model, in the second instance, demonstrated a particularly strong ability to perform complex multi-step scientific reasoning, with the model having, in the technical interpretation, been able to solve a wide range of complex scientific problems that were, by the standards of the period, beyond the capabilities of the standard open source AI models. The model, in the third instance, demonstrated a particularly strong ability to perform complex multi-step code reasoning, with the model having, in the technical interpretation, been able to solve a wide range of complex coding problems that were, by the standards of the period, beyond the capabilities of the standard open source AI models. The model, in the fourth instance, demonstrated a particularly strong ability to perform complex multi-step logical reasoning, with the model having, in the technical interpretation, been able to solve a wide range of complex logical problems that were, by the standards of the period, beyond the capabilities of the standard open source AI models.
The training methodology of DeepSeek R1 was, in the technical interpretation, a particularly significant development in the open source AI ecosystem. The training methodology, in the first instance, was based on a particular form of reinforcement learning, the Group Relative Policy Optimization, GRPO, a reinforcement learning algorithm that was, in the technical interpretation, a particularly efficient and a particularly stable algorithm. The training methodology, in the second instance, was based on a particular reward model, a reward model that was, in the technical interpretation, designed to encourage the model to produce long, detailed chains of reasoning, and the reward model having, in the technical interpretation, been a particularly important source of the quality of the model. The training methodology, in the third instance, was based on a particular training schedule, a training schedule that was, in the technical interpretation, designed to balance the exploration of new reasoning strategies with the exploitation of known reasoning strategies, and the training schedule having, in the technical interpretation, been a particularly important source of the stability of the training. The training methodology, in the fourth instance, was based on a particular inference strategy, a strategy that was, in the technical interpretation, designed to allow the model to allocate more computation to harder problems and less computation to easier problems, and the inference strategy having, in the technical interpretation, been a particularly important source of the efficiency of the model.
The DeepSeek R1 Zero model, in the technical interpretation, was a particularly notable release. The DeepSeek R1 Zero model was, in the first instance, a model that was trained using pure reinforcement learning, without any supervised fine-tuning, and the model was, in the technical interpretation, a particularly important demonstration of the power of the reinforcement learning approach to reasoning. The DeepSeek R1 Zero model was, in the second instance, a model that exhibited a particular emergent behavior, the so-called Aha moment, in which the model, during the training, would spontaneously develop the ability to allocate more time to harder problems, and the Aha moment was, in the technical interpretation, a particularly important demonstration of the power of the reinforcement learning approach to reasoning. The DeepSeek R1 Zero model was, in the third instance, a model that exhibited a particularly strong performance on the standard reasoning benchmarks, and the performance was, in the technical interpretation, broadly comparable to the DeepSeek R1 model on a wide range of the standard reasoning benchmarks. The DeepSeek R1 Zero model, in sum, was a major part of the DeepSeek R1 release, and the DeepSeek R1 Zero model, in sum, was, in the technical interpretation, one of the principal contributions of the DeepSeek R1 release to the open source AI community.
The strategic interpretation of the DeepSeek moment, in the weeks that followed, was, in the strategic interpretation, a major shift in the global AI conversation. The strategic interpretation, in the first instance, was a shift in the global AI conversation about the cost of frontier AI: if a frontier model could be trained for approximately 5.5 million dollars, then the assumption that the frontier of AI was reserved for the well-capitalised American AI laboratories was, in the immediate aftermath, openly questioned. The strategic interpretation, in the second instance, was a shift in the global AI conversation about the location of the frontier of AI: if a small Chinese lab could train a frontier model, then the assumption that the frontier of AI was, in the long run, going to be located in the United States was, in the immediate aftermath, openly questioned. The strategic interpretation, in the third instance, was a shift in the global AI conversation about the openness of the frontier of AI: if a frontier model could be released as open source, then the assumption that the frontier of AI was, in the long run, going to be controlled by a small number of closed-source American AI laboratories was, in the immediate aftermath, openly questioned. The strategic interpretation, in the fourth instance, was a shift in the global AI conversation about the export controls on AI: if a small Chinese lab could train a frontier model on the second-tier H800 GPUs that were, in the period, the principal AI training GPUs available to Chinese AI laboratories, then the assumption that the export controls on the most advanced AI training GPUs were, in the long run, going to be an effective constraint on the Chinese AI industry was, in the immediate aftermath, openly questioned. The DeepSeek moment, in sum, was, in the strategic interpretation, a major shift in the global AI conversation, and a major shift in the global AI strategy conversation.
Open source strategy and the global AI economics
The open source strategy of DeepSeek, in the years between 2023 and 2026, was, in the open source interpretation, a major part of the DeepSeek story. The strategy was, in the first instance, a strategy of releasing the full weights of the DeepSeek models under a permissive open source license, in most cases the MIT license, a license that allows the unrestricted use, modification, and redistribution of the model, including for commercial purposes. The MIT license was, in the open source interpretation, a major departure from the licensing strategies of the principal closed-source American AI laboratories, including OpenAI, Anthropic, and Google, all of which release their models only as API endpoints and not as downloadable model weights. The MIT license was, in the second instance, a strategy of releasing the full training code, the full data preparation code, and the full evaluation code, all under the same permissive open source license, and the release of the full code was, in the open source interpretation, a major part of the open source strategy of DeepSeek. The MIT license was, in the third instance, a strategy of releasing detailed technical papers describing the architecture, the training methodology, and the evaluation results of each of the DeepSeek models, and the release of the detailed technical papers was, in the open source interpretation, a major part of the open source strategy of DeepSeek.
The open source release of DeepSeek R1, in late January 2025, was, in the open source interpretation, a major event in the open source AI ecosystem. The release was, in the first instance, a major event because the model was, by the standards of the period, one of the most capable open source AI models in the world, and the model was, by the standards of the period, broadly comparable to the closed source American frontier models. The release was, in the second instance, a major event because the model was released under the MIT license, and the MIT license was, in the open source interpretation, a major departure from the licensing strategies of the principal closed source American AI laboratories. The release was, in the third instance, a major event because the model was released with detailed technical documentation, and the documentation was, in the open source interpretation, a major contribution to the open source AI community. The release was, in the fourth instance, a major event because the model was, in the immediate aftermath, integrated into a wide range of open source AI tools, including Hugging Face Transformers, the principal open source library for the use of AI models, Ollama, a popular tool for the local execution of AI models, and LM Studio, a popular tool for the local execution of AI models. The release was, in sum, a major event in the open source AI ecosystem, and the release was, in the open source interpretation, a major contribution to the open source AI community.
The community response to the open source release of DeepSeek R1, in the weeks that followed, was, in the open source interpretation, a major event in the open source AI ecosystem. The model was, in the first instance, downloaded at a very high rate from Hugging Face, the principal platform for the distribution of open source AI models, and the model was, in the first few days after the release, the most downloaded model on the Hugging Face platform. The model was, in the second instance, fine tuned by a large number of community contributors, and the fine tuning was, in the open source interpretation, a major source of the rapid improvement of the model in the weeks that followed. The model was, in the third instance, deployed by a large number of community contributors, and the deployment was, in the open source interpretation, a major source of the rapid adoption of the model by the open source AI community. The model was, in the fourth instance, integrated into a large number of community tools, and the integration was, in the open source interpretation, a major source of the rapid adoption of the model by the open source AI community. The community response to the open source release of DeepSeek R1 was, in sum, a major event in the open source AI ecosystem, and the community response was, in the open source interpretation, a major validation of the open source strategy of DeepSeek.
The enterprise impact of the open source release of DeepSeek R1, in the months that followed, was, in the enterprise interpretation, a major event in the enterprise AI ecosystem. The model, in the first instance, was rapidly adopted by a wide range of enterprise customers, including a wide range of small and medium-sized enterprises, a wide range of large enterprises, a wide range of government customers, and a wide range of academic customers. The enterprise adoption of the model was, in the first instance, driven by the substantially lower cost of the model: the model, in the open source interpretation, could be deployed at a substantially lower cost than the closed source American AI models, and the substantially lower cost was, in the enterprise interpretation, a major source of the enterprise adoption of the model. The enterprise adoption of the model was, in the second instance, driven by the substantially greater control of the model: the model, in the open source interpretation, could be deployed with substantially greater control than the closed source American AI models, and the substantially greater control was, in the enterprise interpretation, a major source of the enterprise adoption of the model, particularly for the enterprise customers in regulated industries. The enterprise adoption of the model was, in the third instance, driven by the substantially greater customizability of the model: the model, in the open source interpretation, could be fine tuned at a substantially lower cost than the closed source American AI models, and the substantially greater customizability was, in the enterprise interpretation, a major source of the enterprise adoption of the model.
The follow-up releases of DeepSeek, in the months and years that followed, were, in the technical interpretation, a major source of the long-term impact of the DeepSeek moment. The first major follow-up release, in March 2025, was the DeepSeek V3.1 model, a model that was, in the technical interpretation, a major improvement over the DeepSeek V3 model in a wide range of dimensions, including a substantially improved context length, a substantially improved reasoning capability, and a substantially improved code generation capability. The second major follow-up release, in September 2025, was the DeepSeek V3.2 model, a model that was, in the technical interpretation, a major improvement over the DeepSeek V3.1 model in a wide range of dimensions, including a substantially improved multimodal capability, a substantially improved tool use capability, and a substantially improved agentic capability. The third major follow-up release, in late 2025, was the DeepSeek R2 model, a model that was, in the technical interpretation, a major improvement over the DeepSeek R1 model in a wide range of dimensions, including a substantially improved reasoning capability, a substantially improved general capability, and a substantially improved cost. The follow-up releases of DeepSeek, in sum, were a major part of the long-term impact of the DeepSeek moment, and the follow-up releases of DeepSeek, in sum, were, in the technical interpretation, a major validation of the open source strategy of DeepSeek.
The reaction of the closed source American AI laboratories, in the weeks and months that followed, was, in the closed source interpretation, a major shift in the closed source AI strategy. The OpenAI reaction, in the first instance, was a major shift in the OpenAI strategy: in the months after the release of DeepSeek R1, OpenAI released a number of new models under more permissive terms, including the OpenAI o3 mini model, a model that was, by the standards of the period, broadly comparable to the DeepSeek R1 model, and the release of the OpenAI o3 mini model was, in the closed source interpretation, a major shift in the OpenAI strategy. The Anthropic reaction, in the second instance, was a major shift in the Anthropic strategy: in the months after the release of DeepSeek R1, Anthropic released a number of new models, including the Claude 3.7 Sonnet model, a model that was, by the standards of the period, broadly comparable to the DeepSeek R1 model, and the release of the Claude 3.7 Sonnet model was, in the closed source interpretation, a major shift in the Anthropic strategy. The Google reaction, in the third instance, was a major shift in the Google strategy: in the months after the release of DeepSeek R1, Google released a number of new models, including the Gemini 2.0 Pro model, a model that was, by the standards of the period, broadly comparable to the DeepSeek R1 model, and the release of the Gemini 2.0 Pro model was, in the closed source interpretation, a major shift in the Google strategy. The closed source American AI laboratories, in sum, responded to the DeepSeek moment with a major shift in the closed source AI strategy, and the closed source American AI laboratories, in sum, were, in the long run, going to have to live with the new reality that the open source AI ecosystem was, in the long run, going to be a major force in the global AI ecosystem.
The US China AI policy implications of the DeepSeek moment, in the months that followed, were, in the policy interpretation, a major shift in the US China AI policy. The Biden administration, in the first instance, had, in the years between 2022 and 2024, imposed a series of increasingly restrictive export controls on the most advanced AI training GPUs, the controls being designed to slow the development of the Chinese AI industry. The Trump administration, in the second instance, in the months after the DeepSeek moment, had, in the immediate aftermath, reaffirmed the export controls, and the Trump administration had, in the immediate aftermath, expanded the export controls to include a wider range of AI training GPUs, a wider range of AI training equipment, and a wider range of AI training software. The expanded export controls, in the third instance, were, in the policy interpretation, a major shift in the US China AI policy, and the expanded export controls were, in the policy interpretation, a major response to the DeepSeek moment. The expanded export controls, however, were, in the policy interpretation, not a complete solution to the DeepSeek moment, and the expanded export controls were, in the policy interpretation, a major source of the ongoing tension between the United States and China in the global AI race. The expanded export controls, in the fourth instance, were, in the policy interpretation, a major source of the Chinese effort to develop a domestic AI training GPU industry, an effort that was, in the years that followed, a major part of the Chinese AI industry, and the Chinese effort was, in the policy interpretation, a major source of the ongoing tension between the United States and China in the global AI race.
The aftermath and the meaning of DeepSeek
The aftermath of the DeepSeek moment, in the months and years that followed, was, in the long term interpretation, a major shift in the global AI ecosystem. The first major shift, in the first instance, was a major shift in the global AI conversation: the DeepSeek moment, in the long term interpretation, was a major shift in the global AI conversation about the cost, the ownership, and the future of frontier AI. The shift, in the long term interpretation, was a major shift in the assumption that the frontier of AI was, in the long run, going to be reserved for the well-capitalised American AI laboratories. The shift, in the long term interpretation, was a major shift in the assumption that the frontier of AI was, in the long run, going to be located in the United States. The shift, in the long term interpretation, was a major shift in the assumption that the frontier of AI was, in the long run, going to be controlled by a small number of closed source American AI laboratories. The shift, in the long term interpretation, was a major shift in the assumption that the export controls on the most advanced AI training GPUs were, in the long run, going to be an effective constraint on the Chinese AI industry. The shift, in the long term interpretation, was a major shift in the global AI conversation, and the shift was, in the long term interpretation, a major change in the way that the global AI conversation was conducted.
The second major shift, in the second instance, was a major shift in the Chinese AI ecosystem: the DeepSeek moment, in the long term interpretation, was a major shift in the Chinese AI ecosystem. The shift, in the first instance, was a major shift in the Chinese AI ecosystem about the openness of the frontier: the DeepSeek moment, in the long term interpretation, was a major shift in the Chinese AI ecosystem about the openness of the frontier of AI, and the shift, in the long term interpretation, was a major shift in the way that the Chinese AI ecosystem approached the open source AI ecosystem. The shift, in the second instance, was a major shift in the Chinese AI ecosystem about the relationship between the closed source Chinese AI laboratories and the open source Chinese AI laboratories: the DeepSeek moment, in the long term interpretation, was a major shift in the Chinese AI ecosystem about the relationship between the closed source Chinese AI laboratories (including Baidu, Alibaba, Tencent, and ByteDance) and the open source Chinese AI laboratories (including DeepSeek, Moonshot AI, Zhipu AI, and 01.AI). The shift, in the long term interpretation, was a major shift in the way that the Chinese AI ecosystem approached the open source AI ecosystem. The shift, in the third instance, was a major shift in the Chinese AI ecosystem about the relationship between the Chinese AI industry and the Chinese government: the DeepSeek moment, in the long term interpretation, was a major shift in the Chinese AI ecosystem about the relationship between the Chinese AI industry and the Chinese government, and the shift, in the long term interpretation, was a major shift in the way that the Chinese government approached the Chinese AI industry. The shift, in the long term interpretation, was a major shift in the Chinese AI ecosystem, and the shift was, in the long term interpretation, a major change in the way that the Chinese AI ecosystem operated.
The third major shift, in the third instance, was a major shift in the geopolitical conversation about AI: the DeepSeek moment, in the long term interpretation, was a major shift in the geopolitical conversation about AI. The shift, in the first instance, was a major shift in the geopolitical conversation about the location of the frontier of AI: the DeepSeek moment, in the long term interpretation, was a major shift in the geopolitical conversation about the location of the frontier of AI, and the shift, in the long term interpretation, was a major shift in the way that the geopolitical conversation about the location of the frontier of AI was conducted. The shift, in the second instance, was a major shift in the geopolitical conversation about the export controls on AI: the DeepSeek moment, in the long term interpretation, was a major shift in the geopolitical conversation about the export controls on AI, and the shift, in the long term interpretation, was a major shift in the way that the geopolitical conversation about the export controls on AI was conducted. The shift, in the third instance, was a major shift in the geopolitical conversation about the US China AI race: the DeepSeek moment, in the long term interpretation, was a major shift in the geopolitical conversation about the US China AI race, and the shift, in the long term interpretation, was a major shift in the way that the geopolitical conversation about the US China AI race was conducted. The shift, in the long term interpretation, was a major shift in the geopolitical conversation about AI, and the shift was, in the long term interpretation, a major change in the way that the geopolitical conversation about AI was conducted.
The fourth major shift, in the fourth instance, was a major shift in the public conversation about AI: the DeepSeek moment, in the long term interpretation, was a major shift in the public conversation about AI. The shift, in the first instance, was a major shift in the public conversation about the openness of the frontier of AI: the DeepSeek moment, in the long term interpretation, was a major shift in the public conversation about the openness of the frontier of AI, and the shift, in the long term interpretation, was a major shift in the way that the public conversation about the openness of the frontier of AI was conducted. The shift, in the second instance, was a major shift in the public conversation about the cost of the frontier of AI: the DeepSeek moment, in the long term interpretation, was a major shift in the public conversation about the cost of the frontier of AI, and the shift, in the long term interpretation, was a major shift in the way that the public conversation about the cost of the frontier of AI was conducted. The shift, in the third instance, was a major shift in the public conversation about the safety of the frontier of AI: the DeepSeek moment, in the long term interpretation, was a major shift in the public conversation about the safety of the frontier of AI, and the shift, in the long term interpretation, was a major shift in the way that the public conversation about the safety of the frontier of AI was conducted. The shift, in the long term interpretation, was a major shift in the public conversation about AI, and the shift was, in the long term interpretation, a major change in the way that the public conversation about AI was conducted.
The fifth major shift, in the fifth instance, was a major shift in the research and academic conversation about AI: the DeepSeek moment, in the long term interpretation, was a major shift in the research and academic conversation about AI. The shift, in the first instance, was a major shift in the research and academic conversation about the architecture of frontier AI models: the DeepSeek moment, in the long term interpretation, was a major shift in the research and academic conversation about the architecture of frontier AI models, and the shift, in the long term interpretation, was a major shift in the way that the research and academic conversation about the architecture of frontier AI models was conducted. The shift, in the second instance, was a major shift in the research and academic conversation about the training of frontier AI models: the DeepSeek moment, in the long term interpretation, was a major shift in the research and academic conversation about the training of frontier AI models, and the shift, in the long term interpretation, was a major shift in the way that the research and academic conversation about the training of frontier AI models was conducted. The shift, in the third instance, was a major shift in the research and academic conversation about the deployment of frontier AI models: the DeepSeek moment, in the long term interpretation, was a major shift in the research and academic conversation about the deployment of frontier AI models, and the shift, in the long term interpretation, was a major shift in the way that the research and academic conversation about the deployment of frontier AI models was conducted. The shift, in the long term interpretation, was a major shift in the research and academic conversation about AI, and the shift was, in the long term interpretation, a major change in the way that the research and academic conversation about AI was conducted.
The DeepSeek moment, in sum, was, in the long term interpretation, a major shift in the global AI ecosystem, and a major shift in the way that the global AI ecosystem is, in the long run, going to operate. The shift, in the long term interpretation, was a major shift in the cost of frontier AI, a major shift in the location of frontier AI, a major shift in the openness of frontier AI, a major shift in the geopolitics of AI, a major shift in the public conversation about AI, and a major shift in the research and academic conversation about AI. The shift, in the long term interpretation, was a major shift in the global AI ecosystem, and the shift was, in the long term interpretation, a major change in the way that the global AI ecosystem operates. The DeepSeek moment, in the long term interpretation, was a Sputnik moment for the American AI industry, a Sputnik moment for the American AI strategy, a Sputnik moment for the American AI policy, and a Sputnik moment for the global AI conversation. The DeepSeek moment, in sum, was, in the long term interpretation, one of the most consequential events in the modern history of artificial intelligence, and one of the most consequential events in the modern history of the relationship between the United States and China in the global AI race.
The Liang Wenfeng meeting with Premier Li Qiang, in the weeks after the DeepSeek moment, was, in the political interpretation, a particularly significant moment. The meeting, in the first instance, was a meeting between Liang Wenfeng and the Premier of the State Council of the People’s Republic of China, Li Qiang, and the meeting was, in the political interpretation, a particular recognition by the Chinese central government of the strategic significance of the DeepSeek moment. The meeting, in the second instance, was a meeting at which the Chinese central government, in the political interpretation, signalled a substantial commitment to the further development of the Chinese AI industry, and the Chinese central government, in the political interpretation, committed a substantial amount of additional resources to the development of the Chinese AI industry. The meeting, in the third instance, was a meeting at which the Chinese central government, in the political interpretation, signalled a particular commitment to the open source Chinese AI ecosystem, and the Chinese central government, in the political interpretation, committed a substantial amount of additional resources to the open source Chinese AI ecosystem. The Liang Wenfeng meeting, in sum, was a major event in the political interpretation of the DeepSeek moment, and the Liang Wenfeng meeting, in sum, was, in the political interpretation, a major validation of the strategic significance of the DeepSeek moment.
The implications of the DeepSeek moment for the closed source Chinese AI laboratories, in the months that followed, were, in the closed source Chinese interpretation, a major event in the closed source Chinese AI ecosystem. The closed source Chinese AI laboratories, in the first instance, responded to the DeepSeek moment with a major shift in the closed source Chinese AI strategy: Baidu, the principal closed source Chinese AI laboratory, released, in the months after the DeepSeek moment, a series of substantially more capable closed source models, including the ERNIE 5.0 model, a model that was, by the standards of the period, broadly comparable to the DeepSeek R1 model, and the release of the ERNIE 5.0 model was, in the closed source Chinese interpretation, a major shift in the closed source Chinese AI strategy. The closed source Chinese AI laboratories, in the second instance, responded to the DeepSeek moment with a major shift in the closed source Chinese AI pricing: Baidu, Alibaba, Tencent, and ByteDance all substantially reduced the prices of their closed source AI models, and the substantial price reductions were, in the closed source Chinese interpretation, a major source of the rapid adoption of the closed source Chinese AI models by the Chinese enterprise AI market. The closed source Chinese AI laboratories, in sum, responded to the DeepSeek moment with a major shift in the closed source Chinese AI strategy, and the closed source Chinese AI laboratories, in sum, were, in the long run, going to have to live with the new reality that the open source Chinese AI ecosystem was, in the long run, going to be a major force in the Chinese AI ecosystem.
The implications of the DeepSeek moment for the open source Chinese AI ecosystem, in the months that followed, were, in the open source Chinese interpretation, a major event in the open source Chinese AI ecosystem. The open source Chinese AI ecosystem, in the first instance, expanded substantially in the months after the DeepSeek moment, with a substantial number of new open source Chinese AI laboratories having been founded in the months after the DeepSeek moment, and the substantial expansion having, in the open source Chinese interpretation, been a major source of the rapid development of the open source Chinese AI ecosystem. The open source Chinese AI ecosystem, in the second instance, became substantially more capable in the months after the DeepSeek moment, with a substantial number of new open source Chinese AI models having been released in the months after the DeepSeek moment, and the substantial improvement having, in the open source Chinese interpretation, been a major source of the rapid development of the open source Chinese AI ecosystem. The open source Chinese AI ecosystem, in the third instance, became substantially more integrated with the broader open source AI ecosystem in the months after the DeepSeek moment, with a substantial number of the new open source Chinese AI models having been released on the principal open source AI platforms, including Hugging Face and GitHub, and the substantial integration having, in the open source Chinese interpretation, been a major source of the rapid development of the open source Chinese AI ecosystem. The open source Chinese AI ecosystem, in sum, expanded substantially in the months after the DeepSeek moment, and the open source Chinese AI ecosystem, in sum, was, in the long term interpretation, one of the principal legacies of the DeepSeek moment.
The geopolitical implications of the DeepSeek moment, in the long term interpretation, are still unfolding, and the long term interpretation of the DeepSeek moment, in the long term interpretation, will continue to evolve in the years to come. The geopolitical implications, in the first instance, are the implications for the US China AI race: the DeepSeek moment, in the long term interpretation, was a major shift in the US China AI race, and the DeepSeek moment, in the long term interpretation, was a major shift in the way that the US China AI race is going to be conducted in the years to come. The geopolitical implications, in the second instance, are the implications for the global AI governance: the DeepSeek moment, in the long term interpretation, was a major shift in the global AI governance, and the DeepSeek moment, in the long term interpretation, was a major shift in the way that the global AI governance is going to be conducted in the years to come. The geopolitical implications, in the third instance, are the implications for the global AI safety: the DeepSeek moment, in the long term interpretation, was a major shift in the global AI safety conversation, and the DeepSeek moment, in the long term interpretation, was a major shift in the way that the global AI safety conversation is going to be conducted in the years to come. The geopolitical implications, in sum, are still unfolding, and the long term interpretation of the DeepSeek moment, in sum, will continue to evolve in the years to come.
References
- DeepSeek — Grokipedia, with citations to the standard scholarly literature.
- Liang Wenfeng — Grokipedia, with citations to the standard scholarly literature.
- High-Flyer — Grokipedia, with citations to the standard scholarly literature.
- DeepSeek-V3 Model Card — Hugging Face, December 2024.
- DeepSeek-R1 Model Card — Hugging Face, January 2025.
- DeepSeek-V3 Technical Report — DeepSeek-AI team, arXiv preprint, December 2024.
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning — DeepSeek-AI team, arXiv preprint, January 2025.
- DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model — DeepSeek-AI team, arXiv preprint, May 2024.
- DeepSeek LLM: Scaling Open-Source Language Models with Longtermism — DeepSeek-AI team, arXiv preprint, January 2024.
- NVIDIA H800 Tensor Core GPU — NVIDIA, 2023.
- US Bureau of Industry and Security, October 17 2023 Export Controls on Advanced Computing Items — US Department of Commerce, October 2023.
- US Bureau of Industry and Security, December 2 2024 Expansion of Export Controls — US Department of Commerce, December 2024.
- Nvidia shares plunge as China AI startup DeepSeek rattles markets — Reuters, 27 January 2025.
- DeepSeek: the Chinese AI lab that shook the markets — Financial Times, January 2025.
- DeepSeek and the AI stock shock — The Economist, 30 January 2025.
- How a small Chinese lab shook the global AI industry — The New York Times, 28 January 2025.
- What DeepSeek R1 means for the open source AI movement — MIT Technology Review, 30 January 2025.
- DeepSeek R1 in the community — Hugging Face Blog, January 2025.
- DeepSeek-V3 GitHub Repository — GitHub, December 2024.
- DeepSeek-R1 GitHub Repository — GitHub, January 2025.
- A Survey of Open-Source Large Language Models — Various authors, arXiv preprint, January 2025.
- DeepSeek and the New China AI Race — Foreign Affairs, February 2025.
- Inside DeepSeek: the Hangzhou lab behind the AI shock — South China Morning Post, February 2025.
