R1 1776 is an open-source reasoning model released by Perplexity AI, designed as a post-trained variant of the DeepSeek-R1 architecture. Its primary objective is to provide unbiased, factually accurate responses by removing censorship filters and regional biases found in the original model, particularly regarding sensitive political and geopolitical topics. The model maintains the high-level reasoning, mathematics, and coding capabilities characteristic of the DeepSeek-R1 family.
Technically, R1 1776 is a Mixture-of-Experts (MoE) model with 671 billion total parameters, of which 37 billion are active during inference. Perplexity developed the model using a curated dataset of approximately 40,000 multilingual prompts covering roughly 300 censored topics. The fine-tuning process utilized the NVIDIA NeMo 2.0 framework to ensure that the removal of content restrictions did not degrade its core logical performance or academic benchmark scores.
The name '1776' serves as a symbolic reference to the year of the United States Declaration of Independence, underscoring the creator's emphasis on freedom of information and open discourse in AI development. In addition to the full-scale model, Perplexity released distilled versions, such as a 70B parameter variant, to facilitate broader accessibility for developers with varying hardware resources.