Perplexity open sources R1 1776, a version of the DeepSeek R1 model that Perplexity CEO Aravind Srinivas says was “post-trained to remove the China censorship”
DeepSeek-R1 is a fully open-weight large language model (LLM) … Unite.AI : DeepSeek's R1: A Useful Reminder Markus Kasanmascheff / WinBuzzer : Perplexity Releases Censorship-Free Version of China's DeepSeek R1 AI Reasoning Model Alexey Shabanov / TestingCatalog : Perplexity AI unlocks R1 1776 for developers with open-source release Rani Molla / Sherwood News : Why Amazon has to “keep reducing costs” Maximilian Schreiner / The Decoder : Perplexity AI removes Chinese censorship from Deepseek R1 Ankita Chakravarti / Moneycontrol : New version of DeepSeek-R1 is here, and it will finally answer all questions without avoiding sensitive topics Paul Hill / Neowin : Perplexity releases open-source R1 1776 reasoning model without China censorship Siddharth Jindal / Analytics India Magazine : Perplexity AI Open-Sources R1 1776 to Remove Censorship from DeepSeek-R1 Bluesky: Ed Zitron / @edzitron.com : And there we have it - the first American DeepSeek reasoning model, proof that A) this can be done and B) that everybody is gonna be able to do it. Why pay through the nose for o1? [embedded post] X: Bill Drexel / @bill_drexel : As the Trump Admin reviews the US AISI's mandate, this would be a great service to institutionalize. For all authoritarian open source models like DeepSeek R1: 1) Score the level of censorship/propaganda 2) Post-train and release uncensored version (quickly) 3) Publicly assess security vulnerabilities of provider platform Aaron Ng / @localghost : Exciting work from Perplexity - maximally truthful outputs from AI are the only way forward. I'm very optimistic on the current environment. Nathan Benaich / @nathanbenaich : you can just do things, part 1776 “R1 1776 is a DeepSeek-R1 reasoning model that has been post-trained by Perplexity AI to remove Chinese Communist Party censorship.” [image] @perplexity_ai : DeepSeek-R1 rivals top reasoning models like o1 and o3-mini. However, its usefulness is limited by its refusal to engage with topics censored by the CCP. We aim to always provide accurate answers, but had to address R1's censorship before using its reasoning capabilities. [image] @perplexity_ai : To keep our model “uncensored” on sensitive topics, we created a diverse, multilingual evaluation set of 1000+ examples. Using human annotators and specially designed LLM judges, we compared frequency of censorship in the original R1 and state-of-the-art LLMs to R1 1776. [image] @perplexity_ai : We also ensured that the model's math and reasoning abilities remained intact after the uncensoring process. Benchmark evaluations showed it performed on par with the base R1 model, indicating that uncensoring had no impact on core reasoning capabilities. [image] Aravind Srinivas / @aravsrinivas : We believe it's important reasoning traces are not subject to censorship. You don't want to apply constrained reasoning, but rather one that's maximally truthful. We named the model “R1 1776” in the spirit of the year America got its independence and the American values of free speech. @aravsrinivas : Announcing our first open-weights model: R1 1776 - a version of DeepSeek R1 that's been post-trained to remove the China censorship and provide unbiased, accurate responses. Here's a graph showing % of Chinese censorship by the model (the lower, the better). [image] LinkedIn: Ipek Ceylan : Lately, I have been exploring Perplexity's Deep Research feature announced recently to explore topics in depth when I am looking to craft …
Context & Ripple Effects
DeepSeek R1 had already exposed a tension in open-weight AI: its reasoning capabilities were broadly accessible, while its responses on politically sensitive subjects followed Chinese government restrictions. Perplexity’s release turns that constraint into a downstream post-training target through an uncensored variant of R1.
The move also builds on coverage framing DeepSeek’s openness as a competitive advantage derived from the wider open-source ecosystem. By publishing its modification rather than keeping it proprietary, Perplexity extends that open-research feedback loop to model behavior as well as model capability.
First-order effects
- Developers can download and deploy Perplexity’s R1 1776 weights rather than rely on the base R1’s China-related response restrictions.
- Perplexity positions itself as a distributor and evaluator of a politically differentiated DeepSeek derivative, claiming its post-training preserves the base model’s math and reasoning performance.
Second-order effects
- The release makes response-policy tuning a more visible dimension of competition among open-weight model distributors: downstream teams can compare derivatives on both capability and permitted subject matter.
- Customers considering DeepSeek-based systems gain another deployment option, while model providers face greater pressure to document how alignment and censorship layers affect outputs.
Third-order effects
- If such derivatives proliferate, a model’s country of origin will be a less complete guide to its deployed behavior; governance scrutiny may increasingly shift to the organizations that fine-tune, evaluate, and distribute weights.
- The episode reinforces a split in open AI between resilience through modifiable weights and harder-to-centralize controls over safety, political restrictions, and provenance.
The trend: Open-weight AI is making model access and post-training policy increasingly geopolitical battlegrounds, not just technical implementation choices.