perplexity-ai/r1-1776

by perplexity

Pricing

Input $1.20 / 1M tokens
Output $2.80 / 1M tokens

R1‑1776 Release by Perplexity


Overview

R1‑1776 is a post-trained variant of the DeepSeek-R1 model, open-sourced by Perplexity on February 18, 2025, with the goal of providing uncensored, high-reasoning language generation. It removes politically sensitive filters—particularly around CCP-related topics—while retaining strong general performance.

  • Release Date: February 18, 2025
  • Base Model: DeepSeek-R1
  • Context Length: Up to 128,000 tokens
  • Model Size: 671B parameters

Features

  • Uncensored answers to politically sensitive queries
  • Preserves DeepSeek-R1 reasoning, math, and factual accuracy
  • Open-source weights (MIT license)
  • Available on Hugging Face, OpenRouter, Sonar, and others

Benchmarks

Metric R1‑1776 DeepSeek-R1 GPT‑4 Turbo
Reasoning (lineage-bench) 1st place ≈ equal Top-tier
CCP Topic Coverage 100 % ~2–3 % ≈ 1 %
Multilingual QA High High Very High

Benchmarks reported from Perplexity blog, community evaluations, and ArtificialAnalysis.ai.


Pricing

Token Type Cost per 1M tokens (OpenRouter)
Input $2.00
Output $8.00

Prices vary by provider; self-hosted use is free.


Use Cases

  • Answering geopolitically sensitive questions
  • Multilingual open-topic QA systems
  • Research, journalism, and political science tools
  • Offline/private deployments with full access

Safety & Stability

  • Open-weight model: no censorship, but not aligned to safety standards
  • Post-trained using ~40k prompt completions on censored topics
  • Transparency and reproducibility prioritized

Limitations

  • No native instruction tuning—uses base DeepSeek alignment
  • Some evaluations showed temporarily reduced reasoning (now resolved)
  • Large size: 671B params may require powerful inference infrastructure

License

R1‑1776 is released under the permissive MIT license. It is free to use, modify, and distribute, with the goal of democratizing access to uncensored, high‑reasoning LLMs.