DeepSeek Says Its New Reasoning-Targeted DeepSeek-R1 AI Mannequin Can Outperform OpenAI’s o1

DeepSeek-R1, a reasoning-focused synthetic intelligence (AI) mannequin by the Chinese agency DeepSeek, was launched on Monday. This is the complete model of the open supply AI mannequin, which arrives two months after its preview model was launched. The open-source AI mannequin is offered to obtain, and may also be used as a plug-and-play utility programming interface (API). The Chinese AI agency claimed that DeepSeek-R1 was capable of outperform OpenAI’s o1 mannequin in a number of benchmarks for arithmetic, coding, and reasoning-based duties.

DeepSeek-R1 AI Models Cost Up to 95 Percent Less than OpenAI’s o1

There are two variants within the newest sequence — DeepSeek-R1 and DeepSeek-R1-Zero. Both have been distilled from one other giant language mannequin (LLM) developed by the the AI agency, dubbed DeepSeek V3. The new AI fashions are based mostly on mixture-of-experts (MoE) structure, the place a number of smaller fashions are paired collectively to enhance the effectivity and capabilities of the bigger mannequin.

The DeepSeek-R1 AI fashions are presently out there to obtain by way of its Hugging Face itemizing. The mannequin comes with an MIT licence that enables each tutorial and business utilization. Those, who don’t intend to run the LLM domestically, can go for the mannequin API as a substitute. The firm introduced the inference pricing of the mannequin, highlighting that these price 90-95 % lower than OpenAI’s o1.

Currently, the DeepSeek-R1 API comes with an enter worth of $0.14 (roughly Rs. 12.10) per million tokens and the output worth is about at $2.19 (roughly Rs. 189.50) per million tokens. In comparability, OpenAI’s o1 API prices $7.5 (roughly Rs. 649) per million enter tokens and $60 (roughly Rs. 5,190) per million output tokens.

Not solely does the DeepSeek-R1 price much less, however the firm additionally claims that it presents larger efficiency than the OpenAI counterpart. Based on inner testing, the AI agency acknowledged that DeepSeek-R1 outperformed o1 within the American Invitational Mathematics Examination (AIME), Math-500, and SWE-bench benchmarks. However, the distinction between the fashions is marginal.

Coming to the post-training, the corporate mentioned that it used reinforcement studying (RL) to the bottom mannequin with none supervised fine-tuning (SFT). This technique, also referred to as pure RL, permits extra freedom to the mannequin when fixing complicated issues utilizing the chain-of-thought (CoT) mechanism. DeepSeek claimed that that is the primary open-source AI undertaking to make use of pure RL to enhance reasoning capabilities.