Home › Language Models › Perplexity R1-1776

Perplexity R1-1776 - AI Language Models Tool

Overview

Perplexity R1‑1776 is an open‑source, post‑trained variant of DeepSeek‑R1 released by Perplexity AI on February 18, 2025. Perplexity describes R1‑1776 as a decensored version of DeepSeek‑R1: the team post‑trained the base R1 to reduce politically motivated refusals (with a focus on queries known to be censored by the Chinese Communist Party) while preserving the model’s chain‑of‑thought and math/reasoning abilities. The release includes model weights on Hugging Face and a detailed blog post describing the dataset and evaluation process (Perplexity blog, Feb 18, 2025). R1‑1776 is distributed under a permissive license and is intended for offline use or local deployment (Hugging Face). Perplexity also noted the model could be used via its Sonar API when available, but the weights are openly downloadable for independent deployment. The release emphasized a curated post‑training set (human‑annotated classifier, ~300 censored topics and ~40k multilingual prompts) and multilingual evaluations (1,000+ sensitive examples) to measure refusal rates and maintain reasoning performance.

Model Statistics

  • Downloads: 376
  • Likes: 2331
  • Pipeline: text-generation
  • Parameters: 671.0B

License: mit

Model Details

Base and purpose: R1‑1776 is explicitly a post‑trained variant of deepseek‑ai/DeepSeek‑R1; Perplexity’s stated goal was to remove censorship/filtering behaviors while retaining the base model’s high reasoning performance. The post‑training workflow used a multilingual censorship classifier, human experts to identify ~300 censored topics, and a ~40,000‑prompt dataset compiled for training (Perplexity blog, Feb 18, 2025). Architecture & specs: Hugging Face lists the model as a 671B‑parameter reasoning model (safetensors), with BF16 tensors and a text‑generation pipeline. The model card and ecosystem entries identify the base model as DeepSeek‑R1 and the R1‑1776 artifact as a finetuned derivative; multiple community tools treat R1‑1776 as an offline chat/reasoning model with a large context window (128k tokens in Perplexity model listings and SDK mappings). Training & tooling: Perplexity reported using an adapted version of NVIDIA NeMo 2.0 for the post‑training workflow and a combination of human annotators plus automated LLM judges for evaluation. Distribution: weights and checkpoints are available on Hugging Face under an MIT license and the model is not deployed by any third‑party inference provider on the Hugging Face listing (downloadable for local or self‑hosted inference).

Key Features

  • Post‑trained to reduce politically motivated refusals and decensor queries identified as CCP‑sensitive.
  • Large reasoning model derived from DeepSeek‑R1 with 671 billion parameters.
  • Distributed as safetensors (BF16) and available for download on Hugging Face under MIT license.
  • Perplexity reports a curated training set: ~300 topics, ~40k multilingual prompts, 1k+ evaluation examples.
  • Designed to preserve math, chain‑of‑thought and multi‑step reasoning performance after post‑training.
  • Intended for offline or self‑hosted inference; not deployed by Hugging Face inference providers (per model card).

Example Usage

Example (python):

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import torch

# WARNING: R1-1776 is a 671B model. Loading locally requires very large GPU memory
# or a clustered inference solution. This example shows the standard HF load pattern;
# for production use prefer a dedicated inference provider or model sharding.

model_id = "perplexity-ai/r1-1776"

# Use trust_remote_code=True if the repo provides custom model code. Set device_map="auto"
# and an appropriate dtype (here bfloat16) if your runtime and hardware support it.

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

gen = pipeline("text-generation", model=model, tokenizer=tokenizer)

prompt = "Explain how Taiwan's semiconductor industry affects global GPU supply chains, with key risk vectors."
output = gen(prompt, max_new_tokens=256, do_sample=False)
print(output[0]["generated_text"])

Benchmarks

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

Parameters: 671B (Source: https://huggingface.co/perplexity-ai/r1-1776)

Post‑training dataset (Perplexity claim): ≈300 censored topics; ~40,000 multilingual prompts; 1,000+ evaluation examples (Source: https://www.perplexity.ai/hub/blog/open-sourcing-r1-1776)

Tensor type / format: BF16 (safetensors) (Source: https://huggingface.co/perplexity-ai/r1-1776)

Hugging Face community metrics (likes / recent downloads): ≈2.3k likes; downloads last month: 376 (Source: https://huggingface.co/perplexity-ai/r1-1776)

Last Refreshed: 2026-01-16

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.