Models

← All models

R1 1776

Perplexity · 2025-02-18 · 684.5B parameters

R1 1776 is Perplexity's post-trained variant of DeepSeek-R1, a 37B-active MoE marketed as a 671B model. Released in February 2025, it was trained to reduce politically motivated refusals and evasions present in the base model. Perplexity fine-tuned it on roughly 40,000 multilingual prompts covering approximately 300 sensitive topics identified by human experts while preserving the original model's math and reasoning performance.

Benchmark scores