DeepSeek explained: the company behind the models and the app

DeepSeek reaches people through several doors: its own assistant, a developer platform, downloadable models and, more recently, an agent application called Harness. Those routes connect different businesses. A person can use technology developed by DeepSeek while buying the service from somebody else. Following who publishes the model, who operates it and who receives payment makes the company much easier to understand.[1]

Its origins are unusual for a prominent AI developer. Reuters’s account of DeepSeek’s formation traces it to a research group established in 2023 by High-Flyer, a Chinese quantitative investment firm. High-Flyer founder Liang Wenfeng led the effort. That history explains the connection with algorithmic investing; it does not make every DeepSeek product an investment tool.[2]

One model can travel through several businesses

The rights attached to a released model help make that distribution possible. DeepSeek’s V4.1 Flash repository and weights carry the MIT licence. It permits using, modifying and distributing the material, including commercially, while retaining the required notices. A business can therefore build a service around its own deployed copy. The licence itself does not require paying DeepSeek a fee for each response. The operator still has to supply computing, maintain the deployment and serve its customers.[4]

DeepSeek’s hosted platform is a different commercial relationship. Its terms identify Hangzhou DeepSeek Artificial Intelligence Co., Ltd. as the operator and describe paid services drawing down a prepaid balance. Developers can incorporate those services into their own applications. The terms also expressly permit using API outputs for model training or distillation, subject to the agreement and applicable law. Distillation uses a model’s responses to help train another model. This gives developers a documented route for building something beyond a front end to the original API.[3]

The same terms make a small but useful distinction: using the API alone does not establish an official partnership or endorsement. A product advertising DeepSeek support may simply be a developer’s own application connected to the service.[3]

Hosting changes which organisation handles the request

Microsoft’s current catalogue provides a concrete example of another distribution route. It lists DeepSeek V4.1 Flash as a Preview model hosted on Azure, with purchasing and support through Microsoft. The underlying model carries DeepSeek’s name, but the customer is selecting an Azure offering. Its preview status also matters: the listed service should be assessed on the capabilities and conditions actually available through that endpoint.[5]

Microsoft’s policy for models sold by Azure says it hosts the service without contacting services operated by the original model provider. The policy also distinguishes deployment geographies and warns that preview features can have additional terms. The useful question for a buyer is therefore precise: which hosted offering is being purchased, and which processing terms apply? The name on the model alone cannot answer it.[6]

The company also publishes some of the machinery

DeepSeek’s releases extend into the software that makes large models practical to run. DeepGEMM contains small, specialised computation routines, including the matrix operations used throughout neural networks. Its current repository also includes fused operations that combine related work inside mixture-of-experts models. For engineers, these are reusable implementation details beneath the visible assistant. A 30 September update links a separate implementation for Huawei Ascend hardware, illustrating why an open technical project can develop through distinct hardware-specific versions.[7]

DeepEP tackles communication: moving information between the processors handling different model experts and combining their results. Its repository links research branches and community forks, including AMD-related work. Those are separately identified implementations, with their own requirements. Together, these projects show another route for DeepSeek’s influence: engineers can study or adapt the supporting software, even when the end product is somebody else’s AI service.[8]

Harness expands the application side

Harness is a further step into the software around a model. DeepSeek describes it as an open-source agent framework built around plugins. Those plugins can extend the application rather than requiring every capability to be built into one fixed interface. The repository labels it a developer preview and warns that changes may break compatibility. This article examines its documentation; it is not a hands-on assessment of how dependably the preview completes work.[9]

Its privacy statement makes the provider distinction particularly clear. Harness can use DeepSeek’s official model service or a custom service configured by the user. In the custom route, requests go directly to that provider; the statement says Harness does not store or process those inputs and outputs on its own servers. The selected provider’s terms still apply. With the official service, DeepSeek can collect session logs including user inputs and tool calls. A locally running interface consequently does not mean the model, or all processing, runs on the local computer.[10]

Reach and revenue are different measures

Fresh financing reports should be read with the same attention to what has actually happened. On 6 October 2026, Reuters reported that DeepSeek was expected to raise more than 80 billion yuan, with sources anticipating an October close. Reuters said DeepSeek had not responded to its request for comment. Those are reported expectations for financing, rather than confirmation that the round had closed or a measure of product sales.[11]

A permissively licensed model can reach customers through many operators. That makes model adoption a useful measure of technical reach, while leaving a separate commercial question: which organisation earned the revenue from serving those customers? Understanding DeepSeek means following both the research it releases and the particular service a customer chooses to buy.[4]

Report an error or suggest a correction