DeepSeek to announce enhanced AI model 'DeepSeek-R2' ahead of schedule by May 2025, and to further discount API usage fees by up to 75%


by Prachatai

DeepSeek, a Chinese startup that attracted a lot of attention when it announced the inference model 'DeepSeek-R1' in January 2025, may bring forward the release of its next-generation model ' DeepSeek-R2 ' from its original schedule, Reuters reported. DeepSeek also announced that it will offer up to 75% discounts on API platform usage fees during off-peak hours.

DeepSeek rushes to launch new AI model as China goes all in | Reuters
https://www.reuters.com/technology/artificial-intelligence/deepseek-rushes-launch-new-ai-model-china-goes-all-2025-02-25/

DeepSeek Offers 75% Discount on its Reasoning Model; New R2 Model to Release Before May
https://analyticsindiamag.com/ai-news-updates/deepseek-offers-75-discount-on-its-reasoning-model-new-r2-model-to-release-before-may/

DeepSeek-R1 has attracted attention for achieving high performance with limited computing resources by utilizing technologies called ' Mixture-of-Experts (MoE)' and ' Multihead Latent Attention (MLA).'

Why is DeepSeek causing such a fuss and what's so great about it?



The Chinese government supports DeepSeek, and at least 13 local governments and 10 state-owned energy companies have adopted its model. Another Chinese company, Alibaba, has also developed inference models QwQ-32B-Preview and QwQ-Max-Preview , which are competing with DeepSeek.

According to Reuters, amid fierce competition between DeepSeek and Alibaba, DeepSeek will bring forward the launch of DeepSeek-R2, which was originally scheduled for early May 2025. Compared to the previous model, DeepSeek-R1, DeepSeek-R2 is said to have improved coding capabilities and improved inference capabilities in languages other than English.

DeepSeek also announced that it would offer discounts on the usage fees for its API platform during the off-peak hours of 4:30 p.m. to 12:30 a.m. every day. Input (cache hit) will be $0.035 (about 5 yen) per million tokens, input (cache miss) will be $0.135 (about 20 yen) per million tokens, and output will be $0.550 (about 82 yen) per million tokens. This represents a 50% discount for DeepSeek-V3 and a 75% discount for DeepSeek-R1.



In addition, DeepSeek announced that it will open source five repositories one after another from early March 2025. This move is a sign that as competition in China's AI industry intensifies, companies are vying to open source their products in order to become the standard.

DeepSeek-R1 development companies open-source their proprietary technologies one after another, accelerating AI learning and inference - GIGAZINE



Reuters said, 'The release of DeepSeek-R2 could be a concern for the US government, which has made AI leadership a national priority, and could further invigorate Chinese authorities and companies' AI development efforts.'

Related Posts:

in Software, Posted by log1i_yk