Hybrid thinking model with toggleable chain-of-thought. MATH benchmark 96.6 and AIME 92.1. The most capable local reasoning model for a dual 40GB GPU setup — beats all previous open reasoning models.
This model requires a specialized High-VRAM environment. Ensure you have the latest CUDA Drivers or Metal Framework installed.
Minimum VRAM: 66GB VRAM Recommended
The DeepSeek R2 model by DeepSeek is a 236B (MoE) parameter architecture optimized for reasoning tasks. It requires approximately 64GB of VRAM to comfortably run locally using a Q4_K_M quantization. Extending the context window up to 131,072 tokens will dynamically allocate further VRAM, meaning high-bandwidth memory hardware is strictly advised.
For running DeepSeek R2 at maximum tokens-per-second, we recommend using LM Studio or Ollama with a GGUF quantization (Q4_K_M or Q6_K). If you are multi-GPU, use vLLM to distribute the layers across your VRAM pool for optimal throughput.
You will need a GPU with at least 66GB of VRAM to run the Q4_K_M quantized version smoothly with a moderate context window.
The simplest method is utilizing Ollama by executing 'ollama run deepseek-r2' directly in your command line. Alternatively, you can search for the model via LM Studio's interface.