DeepSeek – the open reasoning model lab
DeepSeek is a Chinese AI lab, founded in 2023 and based in Hangzhou, that released an open-weight reasoning model called DeepSeek-R1 in January 2025 and claimed it was trained for a fraction of the cost usually associated with frontier systems. The company said the reinforcement-learning stage behind its V3 base cost on the order of $5-6 million in compute, a figure that reverberated through equity markets and forced a public reckoning with the assumption that only firms spending hundreds of millions could reach the reasoning frontier. That claim, and the model card and paper that accompanied it, are the reason DeepSeek matters. What follows is an assessment of what the lab actually contributed, how well its numbers hold up, and where its models belong in a real production stack.
DeepSeek in one paragraph
DeepSeek is backed by High-Flyer, a quantitative-trading firm that had already assembled substantial GPU capacity for its own trading models before pivoting some of that capacity toward general AI research. The lab began shipping open-weight models under the DeepSeek name across 2023 and 2024, iterating through a first-generation language model, a coding-focused variant, and a series of mixture-of-experts base models. None of these attracted much attention outside the open-model community. That changed with R1. Where most Chinese labs at the time were chasing GPT-4-class chat quality, DeepSeek published a reasoning model with visible chain-of-thought, open weights, and a permissive license, arriving only a few months after OpenAI's o1 established the category. The combination of capability, openness and a startling cost claim made it, briefly, the most-discussed model release of the year.
The R1 reasoning model
Reasoning models differ from ordinary chat models in that they are trained to spend inference-time compute generating an internal chain of thought before committing to an answer. OpenAI's o1 introduced the commercial category, and its successors extended it; for the underlying idea and lineage, see our explainer on reasoning models, chain-of-thought and deep-think approaches and the companion piece on the OpenAI o-series lineage. R1's significance is that it was the first credible open-weight entry in that category. Anyone could download the weights, inspect the reasoning traces, and run the model without a per-token contract to a US lab.
Architecture: mixture-of-experts, not dense
DeepSeek-R1 is built on the DeepSeek-V3 base, which is a mixture-of-experts (MoE) model rather than a dense one. In an MoE architecture, the network contains many expert sub-networks but activates only a small subset for any given token, so the parameter count that matters for inference is far smaller than the total. V3's total parameter count runs into the hundreds of billions while the active count per token is a fraction of that – the design that makes the model cheap to serve relative to its capacity. This is a deliberate engineering choice: MoE trades memory footprint and routing complexity for lower compute per token, which suits a lab that wants frontier behaviour without frontier inference bills. Exact active-parameter figures are published in the DeepSeek-V3 technical report, which you should treat as the authoritative source since these numbers vary between checkpoints.
GRPO and the training recipe
The methodological contribution most cited from the R1 work is the use of Group Relative Policy Optimization (GRPO), a reinforcement-learning method that dispenses with the separate value-function critic used in more conventional PPO-style RLHF. Instead of training a critic to estimate the value of each state, GRPO samples a group of candidate outputs for a prompt and computes advantages relative to the group's own reward statistics. That removes a large model from the training loop and simplifies the pipeline, which is part of how DeepSeek kept costs down.
The lab also published an earlier variant it called R1-Zero, trained with reinforcement learning applied more or less directly on top of the base model, without the usual supervised fine-tuning warm-up. R1-Zero developed reasoning behaviour on its own but produced messy, sometimes language-mixed output. The production R1 added a cold-start supervised stage and further RL to clean this up. The honest reading of the R1 paper is that the headline is less a single trick than a demonstration that a relatively lean RL recipe, applied to a strong MoE base, can reproduce reasoning behaviour that had been assumed to require far more elaborate and expensive pipelines.
How it benchmarks
On release, independent evaluators including Artificial Analysis and the LMSYS Chatbot Arena placed R1 in the neighbourhood of OpenAI's o1 on mathematics, coding and graduate-level reasoning tasks, while trailing the strongest closed frontier models on some measures. Those standings shift constantly as new checkpoints and rivals appear, so treat any specific ranking as a snapshot rather than a fixed fact and verify against a current leaderboard. For how to read these numbers critically – what MMLU, GPQA and Arena Elo actually measure and where they mislead – see our guide to AI benchmarks. The defensible summary as of early 2026 is that R1 was genuinely competitive with the first generation of commercial reasoning models on public reasoning benchmarks, and that later closed models from OpenAI, Google DeepMind and Anthropic have generally extended their lead on the hardest tasks since.
DeepSeek V3 and the model family
DeepSeek-V3 is the general-purpose base beneath R1, and a capable chat and coding model in its own right. It is the MoE system whose training the cost claims describe, and DeepSeek has continued to ship revised checkpoints under the V3 name, so the version you download today is not necessarily the one benchmarked at launch. Always check the model card for the checkpoint date.
Alongside the large models, DeepSeek released a set of distilled variants that transfer R1's reasoning behaviour into smaller dense models based on other open families, notably several sizes of Qwen and Llama. These distillations are the practical entry point for most builders: they run on a single GPU or, at the smaller end, on a capable workstation, and they retain a meaningful fraction of the parent's reasoning ability. If you want to run one on your own hardware, our walkthrough on running DeepSeek-R1 models locally covers the quantisation and memory trade-offs in detail. The Qwen-based distillations also invite direct comparison with Alibaba's own line, which we profile in Qwen 3, Alibaba's open model family – a useful reminder that DeepSeek is one of several strong Chinese open-model labs, not a lone outlier.
DeepSeek has also shipped an OCR-oriented variant aimed at document understanding – extracting structured text and layout from images of pages, forms and scanned documents. This is a narrower tool than the reasoning models, but it addresses a genuine bottleneck in agentic and retrieval pipelines, where the first job is often turning a PDF into something a language model can actually read.
The license
DeepSeek's releases carry a genuinely permissive license, permitting commercial use and derivative works with far fewer restrictions than the community licenses attached to some Western open-weight models. This is not a token gesture. For teams that need to fine-tune, redistribute, or embed a model in a product without negotiating usage tiers, the license terms are a substantive part of the appeal. As always, read the exact license shipped with the specific checkpoint you intend to use, because terms can differ between models in the family and can change between releases.
The cost-disruption thesis
The number that made DeepSeek famous was the roughly $5-6 million figure attached to the final training run behind V3. Placed against the widely reported figures for training runs at OpenAI, Anthropic, Google DeepMind and Meta – which run into the tens or hundreds of millions of dollars in compute alone – the contrast was stark enough to wipe a large amount of value off chip and AI-adjacent equities in the days after R1's release. The implicit argument was that if a comparatively small lab could reach the reasoning frontier this cheaply, then raw capital spend was a weaker moat than the market had assumed.
That thesis contains a real insight and a real distortion, and separating them matters.
The insight is that algorithmic and architectural efficiency compounds. MoE sparsity, a lean GRPO-based RL recipe, careful engineering around the memory bottleneck, and aggressive use of lower-precision arithmetic together produce a model that is genuinely cheaper to train and to serve than a naive dense approach would be. DeepSeek demonstrated that these efficiencies, stacked, are large. That erodes the crude version of the "scale equals moat" argument.
The distortion is in what the $5-6 million figure includes. By the lab's own framing, it describes the compute cost of the final successful training run – not the cost of the GPUs themselves, not the salaries, not the failed runs and ablations and research iterations that preceded it, and not the substantial prior investment High-Flyer had already made in its cluster. Analysts at outlets such as SemiAnalysis argued that the fully loaded cost of building DeepSeek's capability was very much higher, once hardware capital expenditure and research overhead are counted. Both statements can be true at once: the marginal cost of the final run was remarkably low, and the total cost of standing up a lab able to execute that run was not. Anyone quoting the $5 million figure should say which of the two they mean, because the difference is the entire argument.
The market's overreaction and its later correction are themselves instructive about how thinly the true economics of frontier training are understood, even by sophisticated investors. For the broader picture of how AI costs are reshaping the industry, our coverage of AI economic impact and labour-market analysis puts the DeepSeek episode in context.
Running DeepSeek
There are three routes to running DeepSeek models, each with a different trade-off profile.
Self-hosting the open weights gives you full control, no per-token cost beyond your own compute, and no dependence on any external provider's uptime or data policy. The large MoE models demand serious GPU memory and are realistic only for teams with multi-GPU servers, but the distilled variants bring the smaller end within reach of a single high-memory GPU or even a workstation. This is the route to choose when data residency or air-gapping is non-negotiable.
The second route is a Western inference provider. Companies including Together AI, Fireworks AI and Hyperbolic host DeepSeek models on infrastructure located outside Chinese jurisdiction, exposing them through OpenAI-compatible APIs. This is the pragmatic middle path: you get DeepSeek's capability and cost profile with the operational convenience of a managed endpoint, and your traffic does not traverse Chinese networks. Pricing and available checkpoints differ between providers and change frequently, so compare current rates directly. Managing rate limits and error handling across these providers follows the same patterns as any large-model API; our note on handling 429 rate-limit errors applies broadly.
The third route is DeepSeek's own hosted API, which is typically the cheapest and most current. The caveat is jurisdictional: data sent to DeepSeek's first-party API is processed under Chinese law, and DeepSeek's own terms and privacy documentation should be read carefully before any regulated or sensitive workload touches that endpoint. For many use cases the price is attractive; for others the jurisdiction alone is disqualifying.
The geopolitical context
DeepSeek operates under United States export controls that restrict the sale of the most advanced AI accelerators to Chinese entities. The efficiency story cannot be separated from this constraint: a lab that cannot freely buy the latest chips has a direct incentive to extract more capability per unit of compute, and DeepSeek's engineering choices read partly as a response to scarcity. Whether the lab trained entirely on export-compliant hardware, on chips acquired before the controls tightened, or on some combination has been the subject of considerable reporting and speculation, and the public record does not settle it definitively. Readers should treat categorical claims in either direction with caution.
The release reframed the open-versus-closed debate. Before R1, the strongest open-weight models trailed the closed frontier by a wide margin on reasoning, and the assumption was that the gap would widen as training costs climbed. DeepSeek narrowed the gap from an unexpected direction and did so with weights anyone could download. That is a genuine gain for the open ecosystem, and for researchers and builders who cannot or will not depend on closed APIs.
It also raises legitimate governance concerns that deserve honest treatment rather than either dismissal or alarmism. DeepSeek's first-party hosted service applies content controls consistent with Chinese regulatory requirements, and the model's outputs on politically sensitive topics reflect that. This is a property of the hosted service and of the training data and alignment choices, not an inescapable property of the weights – a self-hosted deployment can be fine-tuned and its guardrails reshaped, though doing so responsibly is non-trivial. The reasonable position is that the open weights are a real technical asset, that the hosted service carries the data-jurisdiction and content-control caveats already noted, and that organisations should decide based on their own threat model rather than on either boosterism or reflexive suspicion. Where model provenance intersects with security, our coverage of AI cybersecurity threats and adversarial AI is relevant background.
Where DeepSeek fits for builders
For cost-controlled reasoning workloads, DeepSeek is one of the strongest available options. If your application needs multi-step reasoning – code generation with verification, mathematical or analytical tasks, agentic planning – and you are sensitive to per-token cost, R1 and its distillations deliver reasoning quality at a price point that closed frontier models struggle to match. Serving from a Western provider or self-hosting keeps the data-jurisdiction question out of the equation.
For open-source compliance scenarios, the permissive license is decisive. Teams building products that must ship model weights, run offline, or avoid any dependence on a single vendor's API terms have few options at this capability level, and DeepSeek is among the best of them. The distilled models in particular make it feasible to embed reasoning capability into constrained environments.
There are places you should not reach for DeepSeek, and they are worth stating plainly. Do not route regulated, personal or sensitive data through DeepSeek's first-party API, where processing falls under Chinese jurisdiction – use a Western-hosted endpoint or self-host instead. Be cautious using any DeepSeek deployment where politically sensitive content handling matters to your users, since alignment choices in the released models reflect their origin. And do not assume DeepSeek is automatically the right choice merely because it is cheap; on the hardest frontier tasks, and for applications that need the strongest available tool-use and long-context behaviour, the leading closed models from OpenAI, Anthropic and Google DeepMind may still justify their higher cost. The disciplined approach is to benchmark the specific models on your own tasks rather than trusting either the marketing or the leaderboard headline.
DeepSeek's lasting contribution is less any single model than a proof: that the reasoning frontier is reachable with disciplined efficiency and released openly, and that the cost structures the industry treated as fixed were more negotiable than they appeared. That proof, more than the weights themselves, is what makes the lab one of the most consequential entrants of the past two years, and why it remains a serious player to watch through 2026.