DeepSeek R1 Models, Distillation and Local Use Explained
DeepSeek published R1, R1-Zero and six smaller distilled checkpoints based on Qwen and Llama model families, with code and weights under documented licenses.
Timeline
- January 2025: DeepSeek published the original R1 repository, model information and weights.
- January 2025: The project listed six distilled checkpoints from 1.5B to 70B parameters.
- After release: The repository and serving projects continued to update their compatibility guidance.
DeepSeek R1 is a family of reasoning models rather than one download suited to every computer. The official repository introduced DeepSeek-R1-Zero, DeepSeek-R1 and six dense distilled checkpoints. The full R1 model uses a mixture-of-experts architecture derived from DeepSeek-V3, while the distilled models use smaller Qwen and Llama base families. [1][2]
R1-Zero was trained with large-scale reinforcement learning without supervised fine-tuning as a preliminary stage. DeepSeek reported that it developed reasoning behaviors but also suffered from repetition, poor readability and language mixing. R1 added cold-start data and further training stages intended to improve usability and alignment with human preferences. [1]
The published distilled checkpoints are Qwen-based 1.5B, 7B, 14B and 32B models and Llama-based 8B and 70B models. Distillation here means fine-tuning smaller base models on data produced from the larger R1 system. A distilled checkpoint is not simply the full model compressed into a smaller file, and its capabilities differ by size and task. [1]
The original repository lists the full R1 and R1-Zero systems at 671 billion total parameters with 37 billion activated for a token. That scale places the full model far beyond an ordinary laptop deployment. Smaller distilled checkpoints are the practical starting point for many local users, but memory, quantization, context length and speed still vary substantially. [1][2]
DeepSeek’s documentation gives serving examples for vLLM and SGLang and instructs users to retain the recommended chat template and generation settings. A command copied from an older README may no longer match the current library release. Users should check the selected checkpoint’s model card and the serving project’s current compatibility documentation. [1][2]
The repository and model weights are published under stated licenses, while distilled models also inherit conditions from their Qwen or Llama base families. “Open weights” does not erase the obligation to review each relevant license. Commercial use, redistribution and hosted services can involve different requirements depending on the checkpoint and jurisdiction. [1]
Running locally can keep prompts on hardware a user controls, but privacy is not automatic. Download sources, extensions, model-serving interfaces and network settings still matter. The model can also produce inaccurate or unsafe output. Verify consequential answers, restrict exposed servers and choose a checkpoint that fits the available memory instead of assuming the largest name will run well. [1][2]