How to Run DeepSeek R1 Locally: Official Starting Points
Running DeepSeek-R1 locally begins with choosing the correct checkpoint. The full mixture-of-experts model has demanding hardware needs, while distilled Qwen- and Llama-based checkpoints offer smaller options that official documentation supports through common serving frameworks.
Timeline
- Choose a checkpoint: Compare the full R1 model with the official 1.5B, 7B, 8B, 14B, 32B and 70B distilled releases.
- Choose a runtime: Confirm that the current framework supports the chosen architecture, precision and hardware.
- Validate locally: Test a small prompt, monitor memory and verify important outputs before integrating the server into another application.
The first step in running DeepSeek-R1 locally is choosing which “R1” you mean. DeepSeek’s full R1 checkpoint is a 671-billion-parameter mixture-of-experts model with 37 billion parameters activated for a token, making it impractical for most personal computers. The same release includes official distilled models at 1.5B, 7B, 8B, 14B, 32B and 70B sizes. Those smaller checkpoints are usually the realistic local starting points. [1][2]
A distilled model is not the full R1 compressed into an identical package. DeepSeek says the smaller checkpoints start from Qwen or Llama base models and are fine-tuned using samples generated by R1. They trade scale and some capability for lower storage and memory demands. Before downloading, read the exact model card, license and tokenizer configuration for that checkpoint; instructions for one size or base family may not transfer safely to another. [1][2]
Hardware requirements depend on parameter count, numerical precision, quantization, context length and runtime overhead. Model weight size is only part of the calculation because inference also needs memory for caches and framework operations. A longer context can substantially increase memory use. Check free disk space and available RAM or GPU memory first, then begin with a smaller official distillation instead of assuming a large checkpoint will fit because its download completed. [1][2]
DeepSeek’s repository provides example serving commands for the distilled Qwen checkpoints using vLLM and SGLang. The current Hugging Face model page also exposes framework-specific starting instructions and an OpenAI-compatible local endpoint pattern. Treat examples as starting points rather than timeless copy-and-paste recipes: verify the framework’s current installation guide, pin versions for repeatability and keep the local server bound to a trusted interface unless remote access is intentionally secured. [1][2]
After the server starts, send a short test request and watch logs for out-of-memory errors, unsupported kernels or tokenizer warnings. Confirm that the response uses the expected model and that generation stops normally. DeepSeek’s usage notes recommend a temperature between 0.5 and 0.7, with 0.6 as the stated default, and advise placing instructions in the user prompt. These are model-author recommendations, not universal requirements for every application. [1][2]
Local execution improves control over where prompts are processed, but it does not automatically create a secure or private deployment. Model files, chat logs, shell history, browser interfaces and network endpoints can still expose data. Download checkpoints from the official DeepSeek organization, review file provenance, restrict server access, avoid feeding secrets during initial testing and update the runtime when security fixes are released. Organizations should apply their own data-handling and software-review policies. [1][2]
A sensible path is therefore: select the smallest official distillation likely to meet the task, confirm its model card and license, install a supported runtime in an isolated environment, start it only on localhost, test memory and output quality, and expand context or model size gradually. Recheck the official repository before deployment because compatibility guidance can change. Regardless of where the model runs, verify consequential code, calculations and factual claims against authoritative sources. [1][2]
Sources
- DeepSeek — DeepSeek-R1 official repository and local-running guidance
- DeepSeek — DeepSeek-R1 official model card on Hugging Face