DeepSeek-R1 & V3
Open-weight frontier reasoning and 671B Mixture-of-Experts architecture
Models
In plain English
The groundbreaking open reasoning model that thinks step-by-step with <think> tags, rivaling OpenAI o1 while running locally on distilled models or 384GB+ Mac/Spark setups.
Technically
A 671B parameter Mixture-of-Experts (MoE) reasoning model by DeepSeek activating 37B params per token (256 routed + 1 shared expert) using Multi-Head Latent Attention (MLA). Distilled models (1.5B–70B on Qwen/Llama) bring reasoning capabilities to consumer GPUs.