Definition
DeepSeek is a family of models developed by DeepSeek-AI. It includes general language models, coding models, and reasoning releases such as DeepSeek-R1. DeepSeek also names the company and its hosted products. An evaluation needs the specific model and deployment.
The family includes models released as downloadable weights and models served through APIs. A hosted model name can be an alias whose underlying release changes. Access to one checkpoint does not establish the weights or license of another.
Origin and attribution
DeepSeek-AI released its early DeepSeek LLM models in November 2023. The team's January 5, 2024 technical report describes the base and chat variants. The repository and report credit the research organization.
DeepSeek released R1 on January 20, 2025. Its report and repository distinguish R1-Zero, R1, and smaller distilled models. R1's report documents reasoning training with verifiable rewards. That training method must be checked separately for other family members.
Scope, lineage, and licenses
The early DeepSeek LLM repository licenses code under MIT while placing model use under a separate model license. The R1 repository applies MIT to its released code and weights, while identifying Qwen and Llama source models for the distilled variants. The upstream lineage still belongs in a release's artifact and license review. One repository badge cannot settle rights across the family.
A distilled model uses training signals from another model. A name such as DeepSeek-R1-Distill-Llama identifies a Llama-derived student, not a smaller copy of R1's original neural architecture.
The current API guide documents compatibility with OpenAI and Anthropic interface formats and names that are served by newer releases. Interface compatibility says how software calls the service; it does not establish identical behavior or output guarantees.
Operational significance
Retain the exact checkpoint or API identifier, reasoning mode, prompt template, and serving configuration. Test the deployment on representative tasks and preserve failed cases. Results from R1, a distilled student, and a later hosted Flash model cannot be substituted without evaluation.
Self-hosting shifts serving and artifact maintenance to the operator. A hosted service shifts some of those duties to its provider while leaving application permissions and verification with the factory. Neither arrangement makes generated answers automatically correct.
Distinguish it from nearby terms
- A reasoning model describes inference behavior; DeepSeek names a family that includes several kinds of models.
- RLVR identifies a training reward source. Its use in a documented release cannot be assumed for every model carrying the brand.
- Distillation describes how a student learns from a teacher. It does not erase the student's base-model lineage.
Check your understanding
A team replaces a self-hosted R1 distilled model with a hosted DeepSeek alias. Which model, license, reasoning, serving, and validation assumptions need to be checked again?