Perplexity AI open-sourced r1-1776, a powerful conversational AI model, democratizing access and fostering community innovation.
When DeepSeek-R1 landed in January 2025, it wiped a reported $600 billion off Nvidia's market cap in a single day and shot to No. 1 on the US App Store, ahead of ChatGPT (Britannica Money, DeepSeek; TechCrunch, "DeepSeek reaches No. 1 on US Play Store," 2025). It was brilliant at reasoning and quietly refused to discuss anything the Chinese government finds inconvenient. Weeks later, Perplexity took the open weights, stripped out that censorship, and released the result. They called it r1-1776. More than a year on, it's still one of the clearest case studies in why open weights matter.
Key Takeaways
- r1-1776 is a post-trained, MIT-licensed version of DeepSeek-R1 (671B params), released February 18, 2025 (DeepLearning.AI, The Batch, 2025).
- It answers topics the base model censored: Perplexity reported the original refused ~85% of sensitive prompts; r1-1776 answers them factually.
- Decensoring did not cost reasoning: it matches DeepSeek-R1 on MMLU, DROP, and MATH-500, and edges it on AIME 2024.
- It runs on Hugging Face weights or Perplexity's Sonar API, making it a practical pick for teams that need auditable, self-hostable reasoning.
Close-up of source code on a monitor, illustrating fine-tuning and post-training of a large language model
r1-1776 is a post-trained version of DeepSeek-R1 that Perplexity open-sourced on February 18, 2025, under the permissive MIT license (Hugging Face, perplexity-ai/r1-1776). It keeps DeepSeek-R1's architecture unchanged: a 671-billion-parameter Mixture-of-Experts model with 37B active parameters and a 128K context window. What changed is its willingness to answer.
The name is a tell. "1776" signals free expression. The whole point of the project was to take a top-tier reasoning model and make it answer questions honestly, regardless of which government would rather it didn't.
r1-1776 is a 671B-parameter, MIT-licensed reasoning model that Perplexity built by post-training DeepSeek-R1 to remove state-aligned censorship. Released February 18, 2025, it preserves the base model's architecture and benchmark performance while answering politically sensitive questions factually, and it's distributed through both Hugging Face and Perplexity's Sonar API.
Because the base model had a blind spot by design. DeepSeek-R1, trained in China, declines or deflects on roughly 300 topics the Chinese Communist Party restricts, from Tiananmen Square to Taiwan's independence (The AI Insider, 2025). Ask it a sensitive question and it tends to repeat state talking points or refuse outright. For a model marketed on reasoning, that's a hole you can't paper over with a system prompt.
Perplexity's fix was post-training, not prompt engineering. They measured the gap, too. On an evaluation set of more than 1,000 multilingual prompts touching censored topics, the company reported that DeepSeek-R1 censored or refused about 85% of them, while r1-1776 answered essentially all of them without that bias.
Think about what that means for a developer. A model that's right 97% of the time on math but evasive on history isn't "mostly reliable", it's selectively unreliable in ways you can't always predict. That's the problem r1-1776 set out to close.
Sensitive prompts censored or refused (%) DeepSeek-R1 (base) r1-1776 (Perplexity) ~85% ~0% Source: Perplexity-reported eval, 1,000+ multilingual prompts (2025).
No, and that's the headline result. Across the benchmarks Perplexity published, r1-1776 sits within a few tenths of a point of DeepSeek-R1 on MMLU (90.5 vs 90.8), DROP (90.92 vs 90.95), and MATH-500 (97.2 vs 97.3), and it actually edges the base model on AIME 2024 (80.96 vs 79.8) (Hugging Face, perplexity-ai/r1-1776). The de-biasing left the reasoning intact.
That parity is the whole proof of concept. It's easy to lobotomize a model into compliance and wreck its capabilities in the process. The hard part is changing what a model will discuss without changing how well it thinks.
Benchmark scores: r1-1776 vs DeepSeek-R1 r1-1776 DeepSeek-R1 90.590.8 90.991.0 97.297.3 81.079.8 MMLUDROPMATH-500AIME 2024 Source: Perplexity published benchmarks, 2025
Source: Perplexity, published benchmark table, 2025
With a focused data pipeline, not a sledgehammer. According to Perplexity's release notes, human experts first identified about 300 topics the CCP censors, then built a multilingual classifier to detect prompts touching them (The AI Insider, 2025). That classifier mined real user queries, keeping only opt-in, PII-filtered prompts, to assemble a training set of roughly 40,000 multilingual examples.
For each, they generated factual, chain-of-thought completions and fine-tuned the model using an adapted version of Nvidia's NeMo 2.0 framework. The deliberate constraint throughout: preserve reasoning quality. They validated that the benchmarks held after training rather than assuming they would.
The interesting design choice here is surgical scope. Perplexity didn't retrain the model's worldview, they targeted ~300 specific censored topics with ~40,000 examples and left everything else untouched. That's why the math and reasoning scores barely moved. The lesson for anyone fine-tuning: the narrower and cleaner your intervention, the less collateral damage to capabilities you didn't mean to change.
Developer workstation lit in blue and red with code on screen, evoking self-hosted AI infrastructure
Because the case for open weights is about control, not just cost. In 2025, open-source models accounted for 11% of enterprise LLM API usage, down from 19% a year earlier, with Chinese open-source models at roughly 1% (Menlo Ventures, "2025: The State of Generative AI in the Enterprise," December 2025). Raw market share dipped, but the strategic reasons to run open weights got stronger.
What r1-1776 demonstrates is that open weights are auditable and modifiable. When a model has a bias baked in, you can only fix it if you can see and retrain it. A closed API gives you neither. For teams in regulated work, sovereignty and the ability to inspect a model's behavior aren't nice-to-haves, they're requirements.
What we tell clients at Apex36: the open-versus-closed question usually isn't about benchmark scores, it's about who gets to change the model's behavior. r1-1776 is the clean example. The capability was already there in DeepSeek-R1; what the open weights bought was the right to fix the part that was broken. That option value is the real reason to keep open models in your stack.
Open-source share of enterprise LLM API usage 19% 11% 2024 2025 Source: Menlo Ventures, State of Generative AI in the Enterprise (2025)
Source: Menlo Ventures, 2025
Two ways, depending on whether you want to host it yourself. The weights are on Hugging Face under the MIT license, so you can download, self-host, fine-tune, and deploy r1-1776 commercially with no usage restrictions (Hugging Face, perplexity-ai/r1-1776). If you'd rather not run a 671B model, Perplexity also serves it through its Sonar API.
Self-hosting makes sense when you need data residency, full auditability, or heavy-volume inference you control. The API route makes sense when you want the decensored reasoning without the infrastructure. Either way, treat it like any open model: run your own evals on your actual use case before you trust it. A benchmark on someone else's test set is a starting point, not a guarantee.
So is it the right model for you? If you need transparent, modifiable reasoning and the freedom to inspect what your model will and won't say, r1-1776 remains a strong, well-documented option. If you're after the latest frontier scores, newer reasoning models have since raised the bar, which is exactly why running your own comparison matters.
For help building that comparison, see our [INTERNAL-LINK: framework for choosing between open and closed LLMs → article on model selection for production teams].
Yes. r1-1776 is released under the MIT license, one of the most permissive open-source licenses, so you can use, modify, self-host, and deploy it commercially without fees or usage restrictions. The weights are freely available on Hugging Face, and Perplexity also offers it through the paid Sonar API for teams that prefer hosted access.
r1-1776 is DeepSeek-R1 post-trained by Perplexity to remove state-aligned censorship on roughly 300 CCP-restricted topics. The architecture (671B params, MoE) and reasoning ability are unchanged. Perplexity reported the base model censored about 85% of sensitive prompts, while r1-1776 answers them factually, with benchmark scores within tenths of a point of the original.
No. On Perplexity's published benchmarks, r1-1776 matches DeepSeek-R1 on MMLU (90.5 vs 90.8), DROP, and MATH-500 (97.2 vs 97.3), and slightly beats it on AIME 2024 (80.96 vs 79.8). The targeted fine-tuning changed which topics the model discusses without degrading its math or reasoning performance.
The model weights are hosted on Hugging Face at huggingface.co/perplexity-ai/r1-1776, with documentation and community resources. Because it's a 671-billion-parameter model, self-hosting needs serious GPU memory. For lighter setups, access it through Perplexity's Sonar API instead of running the full weights locally.
The 1776 reference signals free expression and independence, the project's core goal. Perplexity built r1-1776 to answer questions honestly regardless of political restrictions, so the name frames the release as a stand for open, unbiased information access rather than a purely technical model update.
r1-1776 is a small story with a big lesson. Perplexity took a censored open model, surgically removed the censorship, kept the reasoning intact, and gave it all away under an MIT license. Here's the takeaway:
Talk with our experts about building custom AI solutions for your organization.
Continue exploring these related topics
In 2026, the price of a token fell while the size of the AI bill went up. Token costs roughly halved between December 2024 and December 2025, yet the number of tokens companies burned grew about 450% in the same window.
Reddit removes 100,000 bot accounts daily, yet licenses data to the AI firms whose models generate that same synthetic content back onto the platform.
June 30, 2026, Anthropic released Claude Sonnet 5 and pitched it as the model that finally makes cheap, reliable agents the default. The claim underneath the launch is simple: near-Opus intelligence at a Sonnet price.
