R1 1776
R1 1776 is Perplexity’s significant post-training of DeepSeek R1, targeting refusals and evasive responses around politically sensitive Chinese topics while retaining reasoning capabilities. It is an offline language model and does not use the Sonar search subsystem. The original publisher announcement and model repository have since become inaccessible, but the release is corroborated by the publisher’s February 18 communication, distributor records, and preserved model-card copies and weights. The MIT-licensed 671B checkpoint supports local deployment with substantial hardware resources. Perplexity removed the hosted model on August 1, 2025. Claims of being unbiased are publisher positioning rather than an established guarantee.
2025-02-18
671B total
DeepSeek-R1 Mixture-of-Experts Transformer
MIT
Specifications
- Parameters
- 671B total
- Architecture
- DeepSeek-R1 Mixture-of-Experts Transformer
- License
- MIT
- Context Window
- 128,000 tokens
- Type
- text
- Modalities
- text
Benchmark Scores
Advanced Specifications
- Model Family
- R1 1776
- Finetuned From
- DeepSeek R1
- API Access
- Not Available
- Chat Interface
- Not Available
- Variants
- Preserved third-party quantizations
Capabilities & Limitations
- Capabilities
- reasoningmathematicspolitically sensitive question answering
- Known Limitations
- Original publisher repository and announcement are no longer accessible; preserved weights and cards remainPerplexity API model was removed August 1, 2025128K is the documented historical hosted limitPost-training does not establish neutrality or factual correctness
- Notable Use Cases
- local reasoning researchrefusal-behavior evaluationprivate offline assistants