DeepSeek Deep Dive: The Lab Behind the Cheapest Frontier Models
A brand-and-lab deep dive on DeepSeek itself — origin, funding, model family, architecture, and posture. Not a how-to for the chat app or the API.
DeepSeek is the Chinese AI lab whose open-weight releases in late 2024 and early 2025 forced every major frontier lab to publicly explain their pricing. This page is a deep dive on the lab itself — who runs it, where the money comes from, what has actually shipped, and how the architecture works — not a walkthrough of the chat app or API.
The short version: DeepSeek is a research-first spin-off from a Chinese quantitative hedge fund called High-Flyer. It does not appear on standard startup funding databases because it has not taken outside venture capital. Its GPU stockpile and payroll are underwritten by trading profits, which is a significant part of the story of how it prices its models the way it does.
Its flagship model family, DeepSeek-V3, is a 671B-parameter Mixture-of-Experts model that activates only 37B parameters per token, published under an open license with a full technical report on arxiv.org. Its reasoning model, DeepSeek-R1, was released with weights and a paper describing pure reinforcement-learning-from-rules as a route to chain-of-thought performance.
For a business audience, DeepSeek matters for three reasons: it publishes weights, its API pricing is a floor other providers have to respond to, and its data-handling terms differ meaningfully from US-based providers. Each of those has trade-offs that we cover below.
This page is deliberately narrow. If you want the chat-app onboarding or API quickstart, see How to Use DeepSeek. If you want the per-model deep dive on the V3 architecture, see DeepSeek V3 Explained. If you are evaluating regulated-industry risk, see DeepSeek Data Privacy and Security Risks.
What Is DeepSeek?
DeepSeek (full name DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.) is an AI research lab founded in July 2023 and based in Hangzhou, China. It is led by Liang Wenfeng, who is also the co-founder of the Chinese quantitative hedge fund High-Flyer.
The lab describes its focus as foundational research into large language models and their underlying architecture, rather than end-user consumer products. Its two most visible outputs are the DeepSeek-V3 base and chat models and the DeepSeek-R1 reasoning model, both of which shipped with open weights on Hugging Face and full technical reports.
DeepSeek runs a consumer chat product at chat.deepseek.com and a developer API documented at api-docs.deepseek.com. Both use its own models as the backend. The company has not, as of publication, launched enterprise sales, professional services, or a formal support tier.
Deciding whether DeepSeek belongs in your stack? Layer3 Labs can look at your actual workload, cost profile, and data-handling requirements and give you a specific answer instead of a brand-level one.
Book a ConsultationCompany Origin and Funding
DeepSeek was incorporated in July 2023 as a spin-off from High-Flyer Capital Management, a Hangzhou-based quantitative trading firm that Liang Wenfeng co-founded in 2015. High-Flyer had been running its own AI research group since 2019 and had assembled a large GPU cluster for trading research before turning that infrastructure toward general-purpose language modeling.
The lab is not listed on Crunchbase-style venture databases as having raised an external funding round. Public reporting consistently describes it as bankrolled by High-Flyer's trading profits rather than by outside investors. That single fact explains a lot about how DeepSeek operates: it does not have a board of external investors pushing for revenue milestones, and it has not needed to run a sales motion to keep the lights on.
Its GPU footprint was built up in the years before the October 2023 US export controls tightened on H100-class hardware. Independent analyses have estimated its cluster in the range of tens of thousands of Nvidia A100 and H800 GPUs, though the lab has not published a full inventory. The DeepSeek-V3 technical report on arxiv.org discloses that V3 was trained on a cluster of H800 accelerators.
Model Family Lineage
DeepSeek Coder (November 2023): the lab's first widely distributed release, a family of code-specialized models trained on 2T tokens of code and English, published under an open license on Hugging Face and used to bootstrap the lab's reputation among developers.
DeepSeek LLM (late 2023): general-purpose base and chat models at 7B and 67B parameter scales, released with the training paper and weights.
DeepSeek-V2 (May 2024): the first release using Multi-Head Latent Attention and the lab's Mixture-of-Experts recipe at scale — 236B total parameters with 21B active, and the release that first got DeepSeek attention as an API-cost story.
DeepSeek-V3 (December 2024): 671B total parameters, 37B active, trained on 14.8T tokens, published with a detailed technical report on arxiv.org — see our full breakdown at DeepSeek V3 Explained.
DeepSeek-R1 (January 2025): the lab's reasoning model, trained with a reinforcement-learning-from-rules approach on top of a V3-derived base, released with open weights and a paper describing the RL recipe. Distilled variants at 1.5B through 70B were released alongside.
DeepSeek-V3.1 (2025): an incremental update to V3 with improvements to tool use and long-context handling, distributed through the same open-weights channels.
DeepSeek-R2: publicly discussed as the planned successor to R1. Coverage should treat exact release timing and specs as unconfirmed until an official paper or model card appears; see DeepSeek R2 Explained for the current status of what is and is not public.
Architecture: Why DeepSeek Models Are So Cheap
The V3 technical report is unusually specific about the engineering choices that let a 671B-parameter model be trained and served at a fraction of the cost of a same-scale dense frontier model. Four choices do most of the work.
Mixture-of-Experts with sparse activation: V3 routes each token to a small subset of experts. Total parameter count is 671B, but only 37B are active for any given token, which brings both training FLOPs and inference cost closer to a dense 37B model than a dense 671B one.
Multi-Head Latent Attention (MLA): a compressed KV cache representation that reduces the memory footprint of the attention cache at long context lengths. This is what lets a 128K-context model of this scale be served without exotic memory hardware.
FP8 mixed-precision training: DeepSeek reports training V3 in FP8 for the majority of matrix multiplications, with selective higher-precision accumulation. FP8 halves the memory bandwidth of BF16 and materially cuts training time on H800-class hardware.
DualPipe pipeline parallelism and expert parallelism: the report describes a custom pipeline schedule (DualPipe) that overlaps forward and backward passes to hide the communication cost of expert routing across nodes, plus a load-balancing loss designed to spread traffic across experts without an auxiliary router loss.
None of these techniques were invented at DeepSeek, but combining them in one production training run at this scale, and publishing the recipe, is what makes the reported training cost so low. The lab's public figure for V3 training compute is on the order of 2.788M H800-hours, disclosed in the arxiv.org report.
How DeepSeek Was Trained
V3 pre-training used 14.8 trillion tokens of a mixed corpus. The technical report describes an increased proportion of mathematics and code compared to V2, along with expanded multilingual data (Chinese and English are both first-class).
Multi-Token Prediction (MTP) is used as a training objective in addition to standard next-token prediction. Each MTP module predicts an additional future token, which the report credits with improving data efficiency and enabling speculative decoding at inference time.
Post-training for the DeepSeek-V3-Chat variant follows a supervised fine-tuning phase and a reinforcement-learning-from-human-feedback phase, with reasoning data distilled from DeepSeek-R1 mixed into the SFT stage.
DeepSeek-R1 was trained with a distinct recipe documented in its own paper: a first pass of pure reinforcement learning on rule-based rewards (correctness of math and code) starting from a V3-derived base, followed by a cold-start SFT round and further RL. The paper reports that this recipe produced long chain-of-thought behavior without an SFT-on-reasoning-traces step.
Is DeepSeek Chinese?
Yes. DeepSeek is a Chinese company, headquartered in Hangzhou, incorporated under Chinese law, and staffed primarily in mainland China. Its chat and API services are operated from Chinese infrastructure, and its terms of service and privacy policy are governed by Chinese law.
The practical implications for a business user are: user prompts and generated outputs submitted to the hosted chat app or the hosted API traverse Chinese-operated infrastructure and are subject to Chinese data-handling law. This is a materially different posture from a US-hosted or EU-hosted service and should be treated as a distinct compliance question, not as a marketing objection.
Self-hosting the open-weight models on your own infrastructure removes the network path to DeepSeek entirely — the weights themselves do not phone home. Regulated-industry buyers who want the model capability but not the hosted-service data flow generally take this route.
Is DeepSeek Free?
The consumer chat app at chat.deepseek.com is free to use with account registration. There is no paid consumer tier equivalent to ChatGPT Plus at the time of writing.
The developer API is paid, priced per million input and output tokens, with published rates on api-docs.deepseek.com. Historically these rates have sat at the low end of the market and have been the reference point that pressured other providers to cut their own API pricing.
The open-weight models (V3, R1, and their distilled variants) are free in the sense of licensing — you can download and run them on your own infrastructure without a per-token fee to DeepSeek. Compute cost and operational cost still apply. See DeepSeek pricing for the current published API rates.
Is DeepSeek Safe?
There is no single yes-or-no answer, and treating this as a brand question rather than an architecture question tends to produce bad decisions. The honest framing is: which surface, for which data, under which jurisdiction.
The hosted chat app and hosted API traverse Chinese infrastructure and are governed by Chinese data-handling law. That is a different regulatory posture from a US or EU service, and the right question is whether it fits your specific compliance requirement, not whether it is safe in the abstract.
Independent security researchers have published findings on the hosted chat app's client-side behavior and on the safety-alignment properties of the open-weight models. Both are worth reading before a production decision.
For the full breakdown of the specific trade-offs by surface and by risk category, see DeepSeek Data Privacy and Security Risks.
Open-Weights Posture
DeepSeek publishes weights for its flagship models on Hugging Face and mirrors them on GitHub. The V3 model card on Hugging Face lists the release under a permissive license that allows commercial use, and the code repository is MIT-licensed. Readers should always check the current license file in the model card at the time of use, since license terms can be updated between releases.
The R1 release included the base R1 model and a set of distilled variants derived from Qwen and Llama base models. Distilled variants inherit the license constraints of their upstream base model in addition to DeepSeek's own terms, which matters for commercial redistribution.
DeepSeek is meaningfully more open than a closed-weights lab like OpenAI or Anthropic, and comparably open to Chinese peers like Alibaba's Qwen and Z.AI's GLM series. See Best Open-Weight AI Models for the current comparative landscape.
Business Use Guidance
Reach for DeepSeek when: you need frontier-tier reasoning or code output at the lowest available per-token rate; you are running a workload that benefits from self-hosted open weights (data-sovereignty requirement, high volume where API markup is unacceptable, or a need to fine-tune on private data); you want a second-source model to break single-vendor lock-in.
Be cautious when: you are processing regulated data (PHI, financial account data, EU personal data) through the hosted service without a documented compliance path; you require a formal enterprise SLA and named incident-response contact, which DeepSeek does not currently offer; your buyer has a documented policy against China-operated cloud services.
Realistic middle path: many teams use the hosted API for prototyping and internal-tools work, and self-host the same open-weight model for anything customer-facing or data-sensitive. That gets the cost and capability profile of DeepSeek without the hosted-service data path.
How DeepSeek Compares to Other Labs
Versus OpenAI: OpenAI ships a broader product surface (multimodal generation, voice, agents, first-party enterprise tooling) and offers formal enterprise contracts. DeepSeek publishes weights and prices its API materially lower per token. Trade-off is product breadth and vendor support versus openness and cost.
Versus Anthropic: Anthropic invests heavily in safety research and interpretability, and offers strong enterprise contracting including HIPAA-eligible deployments on major clouds. DeepSeek publishes open weights and a full architecture recipe, which Anthropic does not. Trade-off is safety and compliance posture versus openness.
Versus Z.AI (GLM series): both are Chinese labs publishing frontier-tier open-weight models. GLM's line has been organized more around a chat-product surface; DeepSeek's has been organized around research releases. Both should be evaluated head-to-head on the specific benchmark you care about.
Versus Alibaba's Qwen: Qwen is the other major Chinese open-weights family. It ships more model variants at more parameter counts (multimodal, small models, code-specialized). DeepSeek ships fewer, larger models with more detailed technical reports. Use Qwen when you want a specific size or modality; use DeepSeek when you want the best available open-weight frontier reasoning.
Where DeepSeek Fits
High-volume code assistance for internal developer tooling where per-seat commercial IDE subscriptions get expensive at scale.
Batch reasoning workloads (document analysis, structured extraction, evaluation grading) where API cost dominates and the workload is not customer-facing.
Chinese and East Asian language workloads where the training-data mix favors DeepSeek over US-origin frontier models.
Research and prototyping where reproducibility matters and having the weights and the recipe in hand is a hard requirement.
Second-source deployment behind a model router, alongside a US or EU frontier model, to cap per-request cost on low-risk traffic while routing sensitive traffic elsewhere.
What's Next for DeepSeek
The lab has publicly signaled that DeepSeek-R2 is in development as the successor to R1. Specific release date, parameter count, and architecture details should be treated as unconfirmed until an official model card and paper appear.
The lab has not publicly announced a paid enterprise tier, a first-party fine-tuning service, or a professional-services arm. Anything to that effect should be sourced back to an official DeepSeek announcement before it is planned around.
For a business audience, the more important forward-looking question is not what DeepSeek ships next, but how quickly US and EU providers close the price gap in response. That is the durable market impact of the lab, regardless of the next model release.
Frequently Asked Questions
- DeepSeek is an AI research lab based in Hangzhou, China, founded in July 2023 and spun out of the quantitative hedge fund High-Flyer. It publishes open-weight large language models (notably DeepSeek-V3 and DeepSeek-R1) and operates a consumer chat app and a developer API on top of them.
- DeepSeek's flagship models handle general chat, long-context reasoning up to 128K tokens, code generation, and mathematical reasoning. The R1 model is specialized for chain-of-thought reasoning. Capability is broadly comparable to other frontier models on public benchmarks; run your own evaluations before committing to a workload.
- The V3 architecture is a Mixture-of-Experts transformer with 671B total parameters and 37B active per token, using Multi-Head Latent Attention to shrink the KV cache and Multi-Token Prediction as a training objective. The full recipe is published in the DeepSeek-V3 technical report on arxiv.org.
- DeepSeek charges for API usage per million input and output tokens on its developer platform. The consumer chat app is free. The lab has not taken outside venture funding; its underlying infrastructure and payroll are supported by its parent company, the quantitative hedge fund High-Flyer.
- Four choices, all documented in the V3 technical report: sparse Mixture-of-Experts (only 37B of 671B parameters active per token), Multi-Head Latent Attention to shrink the KV cache, FP8 mixed-precision training, and a custom DualPipe pipeline schedule that hides expert-routing communication cost. Combined at scale on H800 hardware, these cut training and inference cost by a large multiple versus a dense model of similar capability.
- The consumer chat app is free because DeepSeek is a research-driven lab that has not built a paid consumer tier and does not depend on consumer revenue — its parent company is a quantitative trading firm. Revenue comes from paid API usage, not the chat app.
- Yes. DeepSeek is headquartered in Hangzhou, China, incorporated under Chinese law, and its hosted chat and API services run on Chinese infrastructure and are subject to Chinese data-handling law. Self-hosted deployments of the open-weight models do not send data to DeepSeek.
- The weights for DeepSeek-V3 and DeepSeek-R1 are published on Hugging Face and GitHub under licenses that permit commercial use. The training data and the exact post-training datasets are not fully published, which is why the more accurate term is 'open-weights' rather than 'open-source' in the strict sense.
- It depends on the surface and the data. The hosted chat app and API run on Chinese infrastructure under Chinese law, which is a different regulatory posture from US or EU providers — that fits some use cases and not others. Self-hosting the open-weight models removes the hosted-service data path entirely. See our full breakdown at /guides/deepseek-data-privacy-security-risks.
- On public benchmarks, DeepSeek-V3 and DeepSeek-R1 sit near the frontier for reasoning, math, and code, and are frequently the best open-weight option in those categories at the time of writing. Whether it is the best model for your workload is a separate question that depends on task, language, latency requirement, and data-handling constraints.
- DeepSeek-V3 was pre-trained on 14.8 trillion tokens with a Multi-Token Prediction objective, then post-trained with supervised fine-tuning and reinforcement learning from human feedback, with reasoning data distilled from R1. DeepSeek-R1 was trained with reinforcement learning on rule-based rewards on top of a V3-derived base. Both training recipes are published on arxiv.org.
Evaluating DeepSeek for a real workload?
Layer3 Labs helps businesses figure out whether DeepSeek (or another open-weight model) is the right fit for a specific workload, and how to deploy it with the right data-handling posture. We look at your actual traffic, cost, and compliance requirements — not the brand.
/ai-workflow-audit