Zhejiang, January 22: DeepSeek R1 is an LLM launched by Chinese AI Lab DeepSeek that offers powerful performance as the OpenAI o1 models. Recently, the China-based DeepSeek firm launched DeepSeek V-3, which was a MoE (mixture-of-experts) language model trained on 671 billion parameters. It outperformed other rival LLMs, including Meta and OpenAI. The DeepSeek R1 large language model comes with a similar MoE architecture that allows it to do coding, math, and general knowledge-based tasks.

Besides solving these tasks,  DeepSeek R1 is hailed as an LLM that reasons various problems, behaving like humans. The AI model was launched as an open-source by DeepSeek AI lab in China for people to utilise powerful artificial intelligence advancements. Despite being open-source, DeepSeek R1 is a powerful language model that brings a lot to the table regarding performance. Stargate Project: Elon Musk Takes Swipe After OpenAI Announces Mega AI Initiative, Says 'They Don't Actually Have the Money'.

What is DeepSeek R1 LLM? What Does It Do?

DeepSeek R1 is praised in the tech community as a state-of-the-art LLM with advanced problem-solving and analytical capabilities. It was launched as open-source and has a dedicated website: chat.deepseek.com. DeepSeek claimed its new 32B and 70B AI models were on par with OpenAI o1 mini. The company said that DeepSeek R-1. The research paper shared that the new model was launched in two variants: DeepSeek-R1-Zero and DeepSeek-R1. 

The DeepSeek-R1-Zero is trained on RL (reinforcement learning) and did not undergo any supervised fine-tuning. It went through multi-stage RL, ensuring the AI model offered powerful reasoning and readability.  The DeepSeek-R1 was built on the foundation provided by the R1-Zero model. Donald Trump Announces ‘The Stargate Project’ Company With USD 500 Billion Investment in AI, AGI Development With Oracle, NVIDIA and OpenAI Partnership.

DeepSeek R1 Benchmarks and Performance

China-based DeepSeek also shared benchmarks comparing with other models such as OpenAI o1-mini, GPT-4o-0513, Claude 3.5-Sonnet-1022, and others. It worked remarkably well compared to the existing AI LLMs, especially in MATH-500 (Pass@1) and GPQA Diamond (Pass@1) scores. It achieved a higher 96.3 percentile rank of human participants, which represented its expert-level coding abilities. 

(The above story first appeared on LatestLY on Jan 22, 2025 12:37 PM IST. For more news and updates on politics, world, sports, entertainment and lifestyle, log on to our website latestly.com).