The Emergence of DeepSeek: A New Wave in AI Innovation

DeepSeek AI models shown as a compact data centre beside a large industrial plant

The artificial intelligence (AI) industry has been energised by the rapid rise of DeepSeek, a Chinese AI startup redefining how models are developed, trained, and shared. Established in 2023 as an offshoot of the hedge fund High-Flyer, DeepSeek has gained global recognition for its approach to AI model design, particularly its commitment to open-source principles and efficient resource utilisation.

Understanding the technical aspects of DeepSeek's models offers invaluable insights into how constraints can drive innovation.

So, what is DeepSeek? It is a Hangzhou-based AI company that builds large language models and releases their weights publicly, so anyone can download, run and adapt them. Its DeepSeek R1 reasoning model, released in January 2025, matched leading US models at a much lower reported cost, and its DeepSeek V4 models, released in 2026, continue that approach.

What is DeepSeek?

DeepSeek AI is both a company and a family of models. The company is based in Hangzhou and was set up in 2023 by the founder of High-Flyer, which remains a major backer. It offers a chatbot app, a paid API for developers, and open-weight model releases on Hugging Face.

Its two best-known model lines are the general-purpose V series and the R series of reasoning models. The table below shows the main releases.

Major DeepSeek model releases
Model Released What changed
DeepSeek-V3 December 2024 671 billion parameters, with 37 billion active for each token
DeepSeek-R1 January 2025 Reasoning model built on V3, released under the MIT licence
DeepSeek-V3.2-Exp September 2025 Introduced DeepSeek Sparse Attention; API prices cut by more than 50%
DeepSeek-V4 (preview) April 2026 V4-Pro and V4-Flash, with a one-million-token context window
DeepSeek-V4-Pro August 2026 Official release of the flagship model

The R1 release had an immediate effect on markets. On 27 January 2025, Nvidia lost around USD 589 billion in market value, the largest single-day loss in US stock market history, as investors questioned how much computing power leading AI really needs. We looked at that spending question in our piece on the AI infrastructure boom.

How DeepSeek's models are built

DeepSeek's efficiency comes from design choices that reduce how much computing each answer needs, rather than from building the largest possible model. Its DeepSeek-V3 technical report describes three main techniques. If the terms are new to you, our explainer on generative AI vs LLMs covers the basics.

1Mixture of experts

Instead of using the whole model for every word, a mixture-of-experts model sends each token to a small set of specialised sub-networks. DeepSeek-V3 has 671 billion parameters in total, but only 37 billion are active for each token. That keeps the cost of running the model closer to that of a much smaller one.

2Multi-head Latent Attention

Multi-head Latent Attention (MLA) compresses the memory a model keeps about earlier text in a conversation. Smaller memory means each request needs less hardware, which lowers the cost of serving the model.

3Sparse attention, from V3.2 onwards

Sparse attention focuses on the most relevant parts of long inputs instead of processing every token against every other token. DeepSeek introduced it with DeepSeek-V3.2-Exp in September 2025, not in R1 or V3. DeepSeek V4 builds on it with a hybrid design that compresses earlier text further; Hugging Face's analysis estimates V4 needs roughly 2% of the attention memory of a common standard design.

Why is DeepSeek so cheap?

In the face of hardware restrictions imposed by US export controls, DeepSeek has developed training methods to optimise performance on limited computational resources. Three factors explain most of the cost gap.

  • FP8 mixed precision trainingDeepSeek-V3 was trained largely in 8-bit floating point rather than the usual 16-bit formats, which the company says both sped up training and reduced GPU memory use.
  • A reported training cost of USD 5.576 millionDeepSeek's report gives 2.788 million H800 GPU hours for DeepSeek-V3's full training, priced at an assumed USD 2 per GPU hour. The report notes this excludes earlier research and experiments, so it is not the total cost of building the model.
  • Low API pricesDeepSeek cut its API prices by more than 50% with V3.2-Exp. Its current Flash model costs USD 0.15 per million input tokens off-peak, and USD 0.30 at peak times.

Low prices per token do not always mean a low bill, because usage grows with every new workflow. We explain how to budget for that in the LLM pricing trap.

Is DeepSeek open source?

DeepSeek's dedication to open releases has propelled its global recognition. DeepSeek R1 was released under the permissive MIT licence, and DeepSeek publishes its model weights on Hugging Face, so businesses can download the models and run them on their own infrastructure. Strictly, these are open-weight models: the trained model is published, but not all of the training data.

DeepSeek also publishes technical reports with each major release. The research behind DeepSeek R1 was published in the peer-reviewed journal Nature in September 2025.

DeepSeek vs ChatGPT

ChatGPT is OpenAI's product, built on closed models that are only available through OpenAI's own apps and API. DeepSeek offers a similar chatbot and API, but also publishes its model weights, so the same model can be hosted elsewhere. When DeepSeek released R1, it reported performance on par with OpenAI's o1 reasoning model at the time.

For a business, the choice is less about which chatbot is cleverer and more about cost, control and where data is processed. We compare those trade-offs, including security, in our guide to AI model alternatives for business.

What UK businesses can learn from DeepSeek

For technology leaders, DeepSeek's success offers practical lessons. Few businesses will train their own models, so the lessons are about how to choose and use them.

  • Efficiency can matter more than sizeThe most capable model is not always the right one. Matching the model to the task often cuts costs without hurting results.
  • Open-weight models widen your optionsModels you can host yourself give more control over where data is processed, at the cost of running the infrastructure.
  • Build for switchingPrices and capabilities change every few months, so systems should be able to move between models without a rebuild, as we explain in our piece on AI vendor lock-in.
  • Check data handling firstBefore using any hosted AI service, check where data is processed and stored, and whether that fits your obligations under UK GDPR.

DeepSeek's rapid rise in the AI space demonstrates how innovation and collaboration can overcome significant resource constraints. Its focus on software-driven efficiency, efficient architectures, and open releases sets a blueprint for what is possible in a resource-limited environment.

For UK businesses and global enterprises alike, DeepSeek shows how agility and creativity can drive competitive advantage. As artificial intelligence continues to evolve, the story of DeepSeek serves as a reminder of how challenges can lead to real progress.

Choosing the right model for your business?

We help UK businesses compare models on cost, performance and data control, then build LLM and generative AI solutions that can switch models as the market moves.

Talk to us about AI models

Frequently asked questions

What is DeepSeek?

DeepSeek is a Chinese AI company, founded in 2023 and backed by the hedge fund High-Flyer. It builds large language models, including DeepSeek R1 and DeepSeek V4, and publishes their weights openly.

Why is DeepSeek so cheap?

Its models use a mixture-of-experts design that activates only part of the model per token, efficient attention, and FP8 training. DeepSeek also prices its API low, cutting prices by more than 50% in September 2025.

Is DeepSeek open source?

DeepSeek publishes its model weights, and DeepSeek R1 uses the permissive MIT licence. Strictly, they are open-weight models, because the full training data is not published.

What is the latest DeepSeek model?

As of November 2026, DeepSeek's API offers DeepSeek-V4-Pro and DeepSeek-V4.1-Flash, both with a one-million-token context window.

What is the difference between DeepSeek and ChatGPT?

ChatGPT runs on OpenAI's closed models. DeepSeek offers a similar chatbot and API but also publishes its model weights, so its models can be hosted elsewhere. Businesses should compare cost, control and data handling.

Sources

Have a project in mind? Let’s get to work.

Let’s chat about how we can help you. Fill in the details and we’ll get back to you as soon we can.