DeepSeek

DeepSeek is a Chinese artificial intelligence company and a family of open-weight language models. The company was founded in Hangzhou in 2023 by Liang Wenfeng, who had built the quantitative fund High-Flyer, and it began publishing models and technical reports in the same year. Its R1 reasoning model, released in January 2025, brought it global attention: the company later reported that the model had been trained at a fraction of the cost of comparable systems, and it published the weights under a permissive licence 1.

Two features made the releases consequential. The weights could be downloaded, inspected, and run outside the company's servers, and the technical reports described efficiency techniques in unusual detail. Both invited comparisons with closed systems backed by far larger capital budgets, and both shaped the debate about how much computing power frontier AI really requires.

This entry covers the company, its model lineage, the published architecture, licensing, pricing, the market reaction, and the criticism. It distinguishes what the company reported from what independent researchers have established 2.

DeepSeek was founded in 2023 by Liang Wenfeng, a Zhejiang University graduate who had spent more than a decade applying machine learning to financial markets. His quantitative fund, High-Flyer, provided the initial capital and the computing cluster; the AI company grew out of that base rather than from a venture-capital round. It is headquartered in Hangzhou, a city that also hosts Alibaba and a dense cluster of Chinese technology firms 3.

The company's early output was specialised. DeepSeek Coder, released in 2023, targeted programming, and DeepSeek LLM was a 67-billion-parameter general model. Both were accompanied by technical reports and downloadable weights, establishing a pattern that later releases followed. The research team was small by frontier standards, a fact that supporters cite as evidence of efficiency and critics cite as a reason to examine the results closely 1.

Its business is correspondingly narrow. DeepSeek sells API access to its models and operates a consumer chat application; it does not offer the broad enterprise platform, cloud marketplace, or productivity suite of its larger rivals. Revenue figures are not published. The company has said it intends to pursue artificial general intelligence, a claim that places it in the same rhetorical category as much better capitalised laboratories 1.

DeepSeek-V2, released in May 2024, established the architecture the company has largely retained. It was a mixture-of-experts model with 236 billion total parameters and 21 billion active for each token, and it introduced multi-head latent attention, a method for compressing the key-value cache that dominates memory use during inference. The company reported that the design cut inference costs sharply compared with its earlier dense model 4.

V3, released in December 2024, scaled the approach to 671 billion total parameters and 37 billion active. The technical report describes training on 14.8 trillion tokens, multi-token prediction during training, and an auxiliary-loss-free method for balancing load across experts. It states that the final run used about 2.79 million H800 GPU-hours, a figure the company costed at roughly 5.6 million US dollars at two dollars per GPU-hour 4.

R1, released in January 2025, applied reinforcement learning to the V3 base model to elicit long reasoning traces. The company also published R1-Zero, trained without a supervised cold start, and a set of distilled models that transferred some of the behaviour to smaller dense models. The R1 report and a peer-reviewed version in Nature describe the method and release the weights under the MIT licence 2.

The core architectural choice is sparsity. A mixture-of-experts layer routes each token to a small subset of experts, so the model can hold hundreds of billions of parameters while performing arithmetic comparable to a much smaller dense network. Multi-head latent attention compresses the key-value cache into a lower-dimensional latent space, reducing the memory traffic that limits long-context serving. Both techniques lower the cost per token, and both make the model harder to analyse than a dense network of the same active size 4.

Training used FP8 mixed precision, a pipeline schedule that overlaps computation and communication, and a load-balancing objective that avoids collapsing tokens onto a few experts. The report presents these as engineering choices that made a 671-billion-parameter run feasible on export-compliant H800 accelerators, rather than as a single algorithmic breakthrough 4.

R1's contribution is behavioural. The team used group relative policy optimisation with rule-based rewards for answer correctness and format, rather than a learned reward model or a large set of human reasoning demonstrations. R1-Zero developed reflection and self-verification without supervised reasoning traces, though its outputs were often hard to read; the final R1 added a cold-start stage and further training. The Nature version reports that the R1 training run cost about 294,000 US dollars, on top of roughly six million for the base model, figures the authors present as evidence of efficiency rather than as a full account of research spending 2 3.

DeepSeek releases weights, inference code, and technical reports. R1's weights are available under the MIT licence, which permits commercial use, modification, and redistribution with minimal conditions, and the Nature paper states this explicitly 2. V3's weights were released under a separate model licence with use conditions, while its inference code was published under a permissive software licence.

The distinction matters. Open weights let researchers inspect a model's behaviour, run it on local hardware, fine-tune it, and build products without sending data to a third-party API. They also let anyone remove the safety fine-tuning the developer applied, which is why open-weight releases are treated differently from open-source software in safety discussions. DeepSeek does not publish its training corpus or the full data pipeline, so the models can be used and studied but not reproduced from scratch 4.

For the company, openness functions as distribution. A model that can be downloaded and served cheaply spreads faster than one available only through a paid API, and the accompanying reports build credibility with researchers. The approach resembles the strategy of other open-weight publishers: give away the artefact, sell the hosted service, and accept that third parties will compete on price 5.

DeepSeek's API prices are quoted per million tokens and have been low relative to comparable closed models. At the R1 launch, the company listed 0.55 US dollars per million input tokens on a cache miss and 2.19 dollars per million output tokens, with discounts for cached input; V3 was cheaper still. The company has repeatedly cut prices as it improved serving efficiency, and its published rates became a reference point for the Chinese market 5.

The cost structure explains part of the gap. A mixture-of-experts model activates 37 billion of 671 billion parameters per token, so inference uses far less arithmetic than a dense model of the same total size. Latent attention reduces the memory needed to hold a long conversation. But serving at scale still requires clusters of high-end accelerators, and open weights shift that cost to whoever deploys the model rather than eliminating it 4.

The training figures are the more contested economics. The 5.6 million dollar estimate covers the final V3 run at an assumed rental price; it excludes earlier research, failed experiments, staff, data acquisition, and the capital cost of the cluster. The company presents the number as a measure of a single training run, not of the total cost of building the model, and critics who treat it as the latter are misreading it. Even on the narrower reading, the figure was low enough to challenge assumptions about the capital required to compete at the frontier 3.

R1's release coincided with a sharp repricing of AI hardware stocks. On 27 January 2025, Nvidia fell about 17 per cent, erasing roughly 589 billion US dollars of market value, the largest one-day loss in the history of the US stock market, according to Bloomberg data reported by The Straits Times. Other semiconductor and data-centre suppliers fell in sympathy, and the Nasdaq dropped sharply 6.

The move was widely interpreted as a reassessment of how much computing power frontier AI requires. If a capable model could be trained for a few million dollars, the argument went, the demand for accelerators might grow more slowly than investors had assumed. That inference was contested: R1 was trained on existing infrastructure, inference demand could rise as models became cheaper, and the largest cloud providers continued to increase capital spending through 2025 6.

The episode illustrated how quickly a technical result becomes a financial narrative. The market reaction was not a peer review of DeepSeek's claims, and it did not establish that training costs had collapsed. It did show that investors had been treating training cost as a strategic variable, and that a credible low-cost release could move valuations without any change in the underlying hardware 7.

Governments responded on data-protection and security grounds. Italy's data-protection authority ordered a block on the processing of Italian users' data in January 2025 and opened an inquiry into the company's practices. Several other governments restricted the app on official devices, and the model's Chinese ownership made it a test case for how democracies treat AI services subject to another jurisdiction's laws 8.

OpenAI told the Financial Times in January 2025 that it had evidence DeepSeek had distilled its models, a claim also made by US officials. Distillation here means training a model on outputs generated by another model. DeepSeek, in its exchanges with Nature's reviewers, denied copying OpenAI's reasoning examples and said its base model had been trained on internet text that inevitably contains AI-generated material. The accusation has not been independently established, and the distinction between learning from public outputs and improperly extracting a model's behaviour is legally and technically unsettled 7.

The company's own account of its methods has been scrutinised as well. Its training-cost figures are company-reported, its benchmarks are selected by the company, and independent evaluations have found uneven performance outside mathematics and coding. None of that establishes misconduct; it establishes that the evidence base is thinner than the market reaction implied 2.

The models share the weaknesses of the class. They hallucinate, they are sensitive to prompt wording, and their reasoning degrades on problems that require tool use, long-horizon planning, or reliable factual recall. Independent testing has found them strong on competition mathematics and code generation and weaker on some general knowledge, multilingual, and safety evaluations 2.

Service reliability has been a recurring problem. The consumer application has struggled under demand, and API capacity has been constrained at peak times. Documentation is good by open-weight standards but thinner than that of the largest cloud providers, and the company's support operation is small relative to its user base 5.

The governance questions are harder. Data is processed under Chinese law, content controls are not fully documented, and users outside China have limited recourse if an answer is wrong or a request is refused. Open weights mitigate some of this, because a third party can host the model under different terms, but only for organisations with the hardware to do so 1.

Four disputes recur. The first concerns cost. DeepSeek's 5.6 million dollar figure is a single-run estimate that excludes research, staff, and infrastructure; the Nature paper's 294,000 dollar figure covers the R1 training run. Both are defensible as stated and misleading when repeated as the total cost of building a frontier model. The company has not published a full audited budget, and outside estimates of its capital spending are necessarily speculative 4 3.

The second concerns distillation. OpenAI and US officials describe DeepSeek's progress as partly derivative; DeepSeek describes its work as independent and its training data as ordinary web text. The dispute turns on evidence that neither side has published in full, and on whether learning from public model outputs should be treated as infringement, fair competition, or a security problem 7.

The third concerns openness. R1 is genuinely open by the standards of open-weight models, but the training data and pipeline are not, so it does not meet the Open Source Initiative's definition of open-source AI. Calling it open source is common and imprecise; calling it closed is unfair 2.

The fourth concerns capability. Company benchmarks show parity with leading systems on selected tasks; independent evaluations show a strong but uneven model. The gap between the two accounts is not evidence of deception. It reflects the difference between a curated evaluation and the messy distribution of real requests, and it is the ordinary condition of AI benchmarking rather than a DeepSeek peculiarity 2.