DeepSeek (R1 and V-series)

DeepSeek-R1 is a reasoning large language model released by the Chinese AI company DeepSeek on 20 January 2025. It was trained to work through problems step by step before answering, using large-scale reinforcement learning, and its weights were published under the MIT licence, so anyone can download, run and build on it. Its release drew attention because it performed close to leading closed reasoning models at a much lower price.

How it works

R1 builds on DeepSeek’s V3 base model. According to DeepSeek’s technical report, reinforcement learning taught the model to produce a long chain of reasoning before its final answer, which helps on maths, coding and logic tasks. DeepSeek also released smaller “distilled” models, trained on R1’s outputs, that run on modest hardware. You can use R1 through DeepSeek’s chat app and API, or host the open weights yourself or through cloud providers.

DeepSeek has since released newer models, including R1-0528 in May 2025, V3.2 in December 2025 and the V4 family in 2026, so check which version a tool or study refers to.

How it’s used in SEO and marketing

• Drafting, summarising and rewriting tasks, like other chat LLMs.
• Self-hosted processing of sensitive text, since the open weights can run inside your own infrastructure.
• Extracting entities, keywords or attributes from text, with the caveats below.
• Monitoring brand mentions: R1-based assistants are another place where your brand can be described, so they belong in an AI visibility check.

Example

In early 2025 MLforSEO ran the same text through three NLP APIs and four chatbots, including DeepSeek R1, asking each for the entities. R1 returned 38 rows (37 unique entities) across eight types, the most of any chatbot, and found all 13 core people, organisations and brands. Ten of its 37 entities did not appear in the text as written: some were harmless acronym expansions, others were inferences the text did not support, such as job titles the author never claimed. Google Cloud’s Natural Language API returned more entities (59 unique) with salience scores and knowledge base links that no chatbot provided.

Limitations and things to watch

• Chatbot output is not fixed: the same prompt can return different entities on another run, which makes analyses hard to reproduce.
• Entities that are not in the source text are hallucinations for any workflow that maps entities back to pages.
• No salience, confidence or knowledge graph IDs, which purpose-built APIs return.
• Check data handling and where the hosted service stores prompts before sending client or personal data; self-hosting avoids this.
• If you use it for extraction, give it a fixed list of entity types, ask for verbatim text only and verify the output.

Learn it on MLforSEO

• Comparing NLP APIs for Entity and Sentiment Analysis (vs Generative AI)
• What Entity Extraction Is, and Where It Fits in SEO
• Entity & Sentiment Analysis APIs: Comparative Review
• The Introduction to Machine Learning for SEO course, where the entity module includes the API versus generative AI comparison