The word AI, short for artificial intelligence, has become so much of a buzzword that it is hard to hear anything else in technical conversations. One common term is LLM, or large language model: the technology behind products such as ChatGPT and Gemini. We have looked at some Artificial Intelligence topics on the site, but DeepSeek's January 2025 reasoning-model launch was particularly exciting. This Chinese AI developer made downloadable model weights and detailed research available alongside its hosted chatbot. Let's look at what that means for everyday users and homelabs.
What Are Large Language Models and How Do They Work?
LLMs learn statistical patterns from large training datasets and generate text a token at a time. Tokens may represent parts of words, whole words or punctuation. A prompt, previous messages and any supplied documents provide the context for the next response. Fluent output can be useful, but fluency does not establish that an answer is true.
Imagine learning to complete sentences from many examples: “The sky is…” often leads to “blue.” A language model works at a much larger scale and can generate detailed explanations, code and dialogue. It does not consult a verified fact database for every sentence, and a convincing citation can still be invented. Check important claims against the actual source.
How Do LLMs Work?
- Pre-training: The model's numerical parameters are adjusted to learn patterns, commonly through predicting tokens. This is a computationally demanding process distinct from running an already trained model.
- Post-training: Supervised examples and reinforcement learning can improve instruction following and reasoning behaviour. The exact stages differ between releases.
- Inference: A deployed model generates a response using its parameters and the available context. Having a conversation does not automatically retrain those parameters or permanently teach the model your preferences.
A neural network is a mathematical system, rather than a detailed simulation of a human brain. Better benchmarks, longer context windows and faster inference are separate properties. A model may improve one while using more memory or taking longer to produce an answer. Choose the trade-off for your task instead of assuming every newer release is faster and more accurate at everything.
What Is DeepSeek?
DeepSeek names the developer and a family of models, not one unchanging model. This introduction focuses on DeepSeek-V3, released in December 2024, and DeepSeek-R1, released in January 2025. The hosted service and subsequent model releases can change; consult the current documentation when selecting an API model or downloading weights.
V3 is a general language model, while R1 was developed with an emphasis on reasoning tasks such as mathematics and coding. DeepSeek's published launch evaluations reported competitive results on selected benchmarks. Those are the developer's measurements under documented test conditions, not a guarantee that every response will match another service or that a small local model performs like the flagship.
The appealing part is choice: use a browser chatbot for convenience, integrate a hosted API into an application, or run a suitable downloadable model locally. These routes have different hardware, privacy, support and operating costs. A friendly chat interface is an application feature; it should not be confused with the model weights.
Which Capabilities Depend on the Model or Application?
- Current information: Training data is not a live news feed. Search or retrieval can supply newer documents to an application, but this does not continuously update the model's weights. Check dates and linked sources in answers about changing information.
- Input types: The V3 and R1 language-model releases discussed here are text models. Do not assume that an image upload, video analysis or voice feature in a particular application is supported by every DeepSeek checkpoint. Check the exact model card and application's processing path.
- Personalisation: Instructions and conversation context can affect a response. Persistent memory depends on the application and its storage choices; it is not an automatic property of all local models.
- Privacy: Hosted chat and API requests involve an external service. Local inference can keep prompts on your machine, provided the surrounding interface, tools and integrations also run locally. “Open weights” alone does not establish where your data goes.
The V3 and R1 Architecture and Training Advances
Both flagship releases use a mixture-of-experts architecture with 671 billion total parameters and 37 billion activated per token. Routing selects experts for each token. Sparse activation reduces the computation needed compared with activating all those parameters, but the inactive weights still need to be available to the inference system. The active count is not the model's total storage or memory requirement.
V3 also uses Multi-head Latent Attention to reduce attention-cache memory through a compressed representation. Its training report describes additional efficiency techniques, including mixed-precision training. These are meaningful engineering contributions, but there is no single percentage saving that applies to every local runtime, hardware configuration and comparison model.
R1-Zero and R1 are different training experiments. R1-Zero explored reinforcement learning without preliminary supervised fine-tuning. R1 used cold-start data and a multi-stage pipeline with supervised and reinforcement-learning stages to improve reasoning and usability. Describing R1 itself as trained entirely through reinforcement learning misses that distinction.
DeepSeek also released six smaller dense models distilled from R1, based on Qwen and Llama models. Distillation transfers behaviour through training examples; it does not compress the full flagship into a file with identical capabilities. The original distilled sizes were 1.5B, 7B, 8B, 14B, 32B and 70B parameters.
Check the Licence for the Exact Download
The R1 repository and flagship weights are MIT licensed. The original V3 repository separates MIT-licensed code from weights governed by its DeepSeek Model License. R1's distilled releases also identify their Qwen or Llama origins and associated licences. Read the particular model's licence before redistribution or commercial deployment; one permissive announcement does not make every file in the wider family subject to the same terms.
Choosing a Local Model for Your Homelab
| Decision | What to check |
|---|---|
| Checkpoint and size | Exact release, base family and parameter count; smaller distilled models have different strengths from the flagship. |
| Quantisation | Weight precision and file size; reduced precision can lower memory needs while affecting quality. |
| Runtime memory | Weights, context cache and other overhead, including concurrent requests; a download size is not a minimum-RAM specification. |
| CPU and GPU placement | Runtime support and available memory; partial CPU offload can change response speed substantially. |
| Task quality | Representative prompts, correct answers and acceptable waiting time, rather than a parameter count alone. |
Start with a modest model that fits your system and test a few tasks you can verify. Ask it to explain a short piece of code, rewrite a paragraph or answer questions from a supplied document. Record the checkpoint, quantisation and context settings with the results. Increase model size only when the improvement matters enough to justify the additional memory and waiting time.
Trying DeepSeek-R1 Locally with Ollama
Install Ollama using its official instructions, then select a model from its library. The first run downloads weights, so it requires connectivity and sufficient disk space; installation and download time depend on your connection and machine.
# Download and run one model that fits your machine.
ollama run deepseek-r1:8b
# A larger alternative, if your available memory is sufficient.
ollama run deepseek-r1:32b
# Inspect the model and its current CPU/GPU placement.
ollama show deepseek-r1:8b
ollama psTags can change over time. For example, Ollama's current deepseek-r1:8b page identifies R1-0528-Qwen3-8B, whereas the January 2025 distilled 8B release was based on Llama. Record what you actually installed rather than treating the short tag as a permanent model identity.
After the weights are downloaded, local inference can work without a hosted model service. Ollama also offers cloud features, so choose local models deliberately. Its FAQ documents OLLAMA_NO_CLOUD=1 as a way to disable Ollama's cloud models and web search after restarting the server. Separately inspect any web interface, document processor or external tool you add. Keep the service on localhost unless you deliberately configure authenticated access; privacy also requires protecting local chat files and backups.
Practical Ways to Use DeepSeek
DeepSeek can be a useful drafting and thinking assistant: organise ideas for a trip, suggest an email, explain unfamiliar code or generate study questions. Provide the audience, constraints and relevant material, then review the result. For research, ask it to work from supplied sources and check that its summary reflects those sources.
Generating a schedule is different from changing your calendar. Drafting code is different from safely executing it. External actions need an application with tools and suitable permissions. Review generated commands before running them and avoid exposing private documents to a hosted service without understanding its data handling.
For a small business, a limited trial with approved, non-sensitive examples is more informative than an unrestricted chatbot rollout. Count incorrect answers and the time spent correcting them. A tool that saves drafting time can still require human review before sending customer advice or publishing technical instructions.
Conclusion: An Exciting Tool Worth Exploring
DeepSeek's accessible model releases make it an exciting option for experimenting with AI, particularly for people who want to learn how local inference works. It can help with writing, coding and learning, and the range of deployment choices is appealing. Give it a try with a task you understand, verify the result and choose the model and privacy arrangement that suit your needs.
Visit the official DeepSeek Website or the DeepSeek Chatbot. For technical details, see the R1 repository, original V3 repository and Ollama FAQ.
