Home › Language Models › Perplexity R1-1776

Perplexity R1-1776 - AI Language Models Tool

Overview

R1‑1776 is a post‑trained variant of DeepSeek‑R1 released by Perplexity AI to provide uncensored, factual answers while retaining strong multi‑step reasoning. Perplexity published the weights and model card on Hugging Face and described a focused post‑training effort that identified ~300 CCP‑censored topics and used a curated 40k multilingual prompt/response corpus to teach the model to answer sensitive queries directly. ([perplexity.ai](https://perplexity.ai/hub/blog/open-sourcing-r1-1776)) The release (published Feb 18, 2025) positions R1‑1776 as an open‑weight reasoning model suitable for local deployment, research, and experimentation; Perplexity also offered access via its Sonar API at launch but later moved away from serving R1‑1776 in production API endpoints. Community reaction has been mixed — with independent benchmarkers and users reporting both regressions and parity with the original R1, and Perplexity publishing internal evaluations showing preserved math and chain‑of‑thought performance. ([perplexity.ai](https://perplexity.ai/hub/blog/open-sourcing-r1-1776))

Model Statistics

  • Downloads: 1,135
  • Likes: 2333
  • Pipeline: text-generation

License: mit

Model Details

Base & lineage: R1‑1776 is a post‑trained finetune of deepseek-ai/DeepSeek‑R1 (DeepSeek family). Perplexity describes the work as a post‑training pass focused on removing politically driven censorship while preserving reasoning. ([huggingface.co](https://huggingface.co/perplexity-ai/r1-1776)) Architecture & size: The Hugging Face model card and community mirrors list the model at ~671B parameters and provided as safetensors with BF16 weights; multiple community quantizations (GGUF/8‑bit etc.) are available for local use. ([huggingface.co](https://huggingface.co/perplexity-ai/r1-1776)) Training & evaluation: Perplexity reports collecting ~40k multilingual prompts targeted at ~300 identified censored topics, then post‑training with an adapted NVIDIA NeMo 2.0 workflow. They evaluated R1‑1776 on a curated set of 1,000+ sensitive examples using human annotators and LLM judges and report parity on math/reasoning benchmarks with the base R1. ([perplexity.ai](https://perplexity.ai/hub/blog/open-sourcing-r1-1776)) Context window & I/O: Community and tooling pages list an extended context capability (128k tokens) and support for long chat completions; R1‑1776 is described as a text‑only reasoning model (no native multimodal inputs). Deployments typically require large‑memory GPUs or quantized runtimes. ([glama.ai](https://glama.ai/models/r1-1776?utm_source=openai)) Distribution & licensing: Perplexity published the model weights on Hugging Face under a permissive license (MIT shown on the model card), enabling research and local deployment. Note: Perplexity deprecated serving R1‑1776 from its hosted API lineup in mid‑2025; the weights remain downloadable. ([huggingface.co](https://huggingface.co/perplexity-ai/r1-1776))

Key Features

  • Post‑trained to reduce CCP‑aligned censorship on identified sensitive topics.
  • Preserved chain‑of‑thought reasoning and math capabilities after post‑training.
  • Curated multilingual dataset of ~40k prompts targeting ~300 censored topics.
  • Large model (~671B) with community quantizations for local deployment.
  • Extended context support reported (≈128k tokens) for long documents and chats.

Example Usage

Example (python):

from huggingface_hub import snapshot_download
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

# 1) Download the model files (large; ensure sufficient disk space)
repo_id = "perplexity-ai/r1-1776"
local_path = snapshot_download(repo_id)

# 2) Tokenizer + model (note: this model is very large and may require special runtimes,
#    quantization, or HF Text‑Generation‑Inference to run efficiently)
tokenizer = AutoTokenizer.from_pretrained(local_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(local_path, trust_remote_code=True, device_map="auto")

gen = pipeline("text-generation", model=model, tokenizer=tokenizer)
resp = gen("Explain how Taiwan's independence could affect global semiconductor supply chains.", max_new_tokens=256)
print(resp[0]["generated_text"])

# Notes:
# - Running the full 671B model locally requires specialized hardware and memory.
# - Many users employ quantized GGUF builds, llama.cpp, or hosted inference services instead.
# - Perplexity initially offered API access via Sonar; R1‑1776 weights remain downloadable on Hugging Face. ([huggingface.co](https://huggingface.co/perplexity-ai/r1-1776))

Benchmarks

Release date: February 18, 2025 (Source: ([perplexity.ai](https://perplexity.ai/hub/blog/open-sourcing-r1-1776)))

Model size: Approximately 671B parameters (Source: ([huggingface.co](https://huggingface.co/perplexity-ai/r1-1776)))

Context window: 128,000 tokens (reported in community/tooling pages) (Source: ([glama.ai](https://glama.ai/models/r1-1776?utm_source=openai)))

Post‑training dataset: ≈40,000 multilingual prompts targeting ~300 censored topics (Source: ([perplexity.ai](https://perplexity.ai/hub/blog/open-sourcing-r1-1776)))

Evaluation set size: 1,000+ curated sensitive examples (human + LLM judges) (Source: ([perplexity.ai](https://perplexity.ai/hub/blog/open-sourcing-r1-1776)))

Hugging Face downloads (last month): 1,135 (as shown on model page at time of capture) (Source: ([huggingface.co](https://huggingface.co/perplexity-ai/r1-1776)))

Last Refreshed: 2026-03-03

Key Information

  • Category: Language Models
  • Type: AI Language Models Tool
Visit Official Website

Related Tools

Qwen2.5-7B

Qwen2.5-7B is a large language model designed for text generation, featuring improvements in coding, mathematics, instruction following, long text generation, and multilingual support. It supports context lengths up to 128K tokens and is intended for sophisticated NLP tasks.
DeepSeek-V3

Open‑weight flagship LLM (base and chat) with strong reasoning and tool‑use; runs with SGLang/LMDeploy/vLLM and permissive model license.
Llama 3

Llama 3 is an open access large language model (LLM) released by Meta, available in various configurations (8B and 70B parameters) with capabilities for fine-tuning and integrations into platforms like Hugging Face, Google Cloud, and Amazon SageMaker.
UNfilteredAI-1B

A large-scale text generation model designed for creative and unconstrained content production without traditional filtering.
Shuttle-3

Shuttle-3 is a state-of-the-art language model designed for high-quality text generation, particularly suited for complex chat, multilingual communication, and reasoning tasks. It is fine-tuned from the Qwen-2.5-72b-Instruct model and designed to emulate high-quality prose similar to Claude 3 models.
WizardLM

WizardLM is a state-of-the-art large language model designed for complex chat, multilingual tasks, reasoning, and agent functionalities. It features an AI-powered pipeline (Auto Evol-Instruct) that optimizes instruction datasets for improved performance across various domains and leverages Arena Learning for an expanded learning pool of challenging instruction data.