DeepSeek describes its current foundational models as large language models based on deep neural networks, with separate training and inference stages. This is its description at the time of the disclosure.
Supporting quotations
Currently, the foundational models provided by DeepSeek are all large-scale language models based on deep neural networks. These models operate in two main stages: the training phase and the inference phase. Additionally, DeepSeek models are open-source, so this section will also introduce our open-source efforts.
Use & Processing · Company practice
Training adjusts model parameters in two steps
DeepSeek describes development of deployable neural-network models with billions to trillions of parameters, optimized through gradient descent. It generally divides training into pre-training and optimization training.
Supporting quotations
The model training phase is the development stage, where developers create deployable models using designed training methods. The models consist of multi-layered neural networks with parameters ranging from billions to trillions. These parameters are continuously optimized during training through gradient descent algorithms. Model training can generally be divided into two steps: pre-training and optimization training.
Use & Processing · Company practice
Pre-training builds language ability, with further adjustments needed
Pre-training uses large-scale self-supervised learning on text to learn language patterns and knowledge associations. DeepSeek says this can produce coherent text, but may not yet yield accurate answers or task performance, so further training adjustments are needed.
Supporting quotations
Pre-training: The goal of pre-training is to train the model with datasets, enabling it to
acquire general language understanding and generation capabilities. During this phase, the model learns language patterns and knowledge associations from text data through large-scale self-supervised learning. After pre-training, the model can understand and generate coherent text but may not yet answer questions or perform tasks accurately. Further training adjustments are required.
Use & Processing · Company practice
Fine-tuning adapts the model to practical tasks
Optimization training, also called fine-tuning, adjusts the pre-trained model with task-specific data. DeepSeek says it typically uses supervised fine-tuning or reinforcement learning to follow instructions, align with human preferences, and develop domain expertise. Better practical performance is not a promise of accurate answers in every case.
Supporting quotations
Optimization Training: Also known as fine-tuning, optimization training further adjusts
the model parameters based on the pre-trained model using task-specific data to adapt it to real-world application scenarios. During this phase, the model typically employs supervised fine-tuning (SFT) or reinforcement learning (RL) methods to learn how to answer questions according to instructions, align with human preferences and needs, and unlock expertise in specific domains. After optimization training, the model better meets practical requirements and can be deployed.
Use & Processing · Company practice
The deployed model generates responses from user input
During inference, DeepSeek models compute over input and predict the next token. In its product services, this produces responses such as text, tables, and code. The disclosure also describes use in text-based tasks and downstream applications.
Supporting quotations
The inference phase is when the model is deployed to provide services. Once trained and deployed, the model can encode and compute input information to predict the next token, thereby enabling text generation, conversation, and other capabilities. It can proficiently perform a wide range of text-based tasks and be integrated into various downstream systems or applications. Specifically, for DeepSeek's product services, the model computes and infers based on user input to generate corresponding responses, including text, tables, and code.
Use & Processing · Company practice
DeepSeek describes probabilistic generation rather than retrieval of training copies
DeepSeek says its model predicts token sequences probabilistically from the input context, rather than simply retrieving or copying training text. It states that the model does not store copies of the original training data. This describes the model, not a promise that the company retains no training or user data elsewhere.
Supporting quotations
It is important to note that the model uses an autoregressive generation method, predicting the most likely subsequent sequence of tokens based on the input context through probabilistic calculations. This process is not a simple retrieval or "copy-paste" of text from the original training data. The model does not store copies of the original training data but dynamically generates contextually appropriate responses based on its deep understanding of language structure and semantic relationships.
Use & Processing · Restriction or exception
DeepSeek cannot guarantee that model outputs are accurate
DeepSeek warns that AI may produce incorrect, omitted, or non-factual content, described as hallucination. It commits to reducing hallucinations through research, including higher-quality training sources, alignment strategies, and retrieval-augmented generation, but expressly cannot guarantee hallucination-free output. It says warning labels appear on the welcome page, at the end of generated text, and at the bottom of the interactive interface.
Supporting quotations
Currently, AI is still in its early stages, and the technology is not yet mature. Due to the limitations of current model principles, AI may generate incorrect, omitted, or non-factual content, a phenomenon known as "hallucination." Hallucination is a challenge faced by the entire AI industry. DeepSeek is committed to reducing hallucination rates through research, including but not limited to selecting high-quality training data sources, optimizing alignment strategies, and employing retrieval-augmented generation (RAG) techniques. However, at this stage, we cannot guarantee that the model will not produce hallucinations. To further mitigate the potential adverse effects caused by hallucinations, we have added prominent warning labels on DeepSeek's welcome page, at the end of generated text, and at the bottom of the interactive interface, specifically reminding users that the content is AI-generated and may be inaccurate.
Cross-references
qualifies (model-proposed): The deployed model generates responses from user input — qualifies: The model’s ability to generate responses does not guarantee correctness: DeepSeek expressly warns of inaccurate, missing, or non-factual output.
Use & Processing · Context
Separate training-data summaries are linked for named model versions
The disclosure links training-data summaries for DeepSeek-V3.1 and V3.2, V4, and V4.1. Those PDFs are not supplied here, so this card identifies the references without claiming what they contain.
The linked training-data summaries’ text is not supplied here; their contents cannot be analyzed from these links.
Use & Processing · Company practice
A small portion of fine-tuning data may be based on user input
DeepSeek says its research team produces question-answer data for optimization training, manually or automatically, with a small portion potentially based on user input. If it uses user input, it applies encryption, strict de-identification, and anonymization to prevent linkage to a specific individual. Users have a right to opt out of AI training, with instructions referred to the separate Privacy Policy.
Supporting quotations
During the optimization training phase, we typically need to construct or annotate a set of question-answer pair data manually or automatically to train the model. These question-answer pairs are produced by our research team, with a small portion potentially based on user input. If user input is used to construct training data, we apply secure encryption, strict de-identification, and anonymization to make it cannot be linked to any specific individual. We also deploy measures to make personal information not appear in the model's outputs to other users and we will not use it for user profiling or personalized recommendations. To be clear, we do not offer services involving user profiling or personalized recommendations. Users are also given the right to opt out. For information on how to opt out of AI training, please refer to the DeepSeek Privacy Policy.
Not stated here
The linked policy’s opt-out procedure is not supplied here.
Use & Processing · User right or choice
Users can opt out of AI training
DeepSeek gives users the right to opt out of AI training and directs them to its separate Privacy Policy for instructions. That policy also supplies specific personal-information rules; the current document does not provide the full opt-out process.
Supporting quotations
Users are also given the right to opt out. For information on how to opt out of AI training, please refer to the DeepSeek Privacy Policy.
For specific rules on how we collect, protect, and use personal information, please carefully read the DeepSeek Privacy Policy.
Not stated here
Opt-out steps, timing, and any further conditions in the linked policy are not supplied.
Cross-references
qualifies (model-proposed): A small portion of fine-tuning data may be based on user input — qualifies: Potential use of a small portion of user input for fine-tuning is accompanied by a right to opt out; the supplied document refers users to the separate Privacy Policy for instructions.