Baidu Announces Two Main Large Models Will Be Available for Free

After the significant price reduction of ByteDance and Alibaba Cloud’s large models, Baidu‘s two main large models are announced to be free.

On May 21st, Baidu AI Cloud announced that its two main models ENIRE Speed and ENIRE Lite of the Baidu ENIRE large model will be completely free, effective immediately.

It is reported that these two large models were released in March this year, supporting 8K and 128K context lengths. Baidu AI Cloud’s official website shows that ERNIE Lite is a self-developed lightweight large language model by Baidu, Inc., balancing excellent model performance and inference capabilities, suitable for low-power AI acceleration card inference use. ERNIE Speed is Baidu‘s self-developed high-performance large language model with outstanding general capabilities, suitable for fine-tuning as a base model to better handle specific scenario problems while possessing excellent inference performance. In addition, Baidu also has a lightweight model called ERNIE Tiny.

According to the previous pricing, ERNIE Lite charges 0.003 yuan per thousand tokens for input and 0.006 yuan per thousand tokens for output. The price of ERNIE Speed is 0.004 yuan per thousand tokens for input and 0.008 yuan per thousand tokens for output.

On the same day, Alibaba Cloud also announced that the API input price of its main model Qwen-Long has been reduced from 0.02 yuan per thousand tokens to 0.0005 yuan per thousand tokens, a direct decrease of 97%.

It is worth noting that, before Baidu and Alibaba, the pricing of DouBao large model under ByteDance was considered to be the beginning of a price war. The main model of DouBao in the enterprise market was priced at only 0.0008 yuan/thousand tokens, which could process more than 1,500 Chinese characters for 0.8 cents, making it 99.3% cheaper than the industry at that time. Also priced in “cents” is Zhipu AI; recently, Zhipu AI announced a promotional policy where new registered users will receive an increase in free quota from 5 million tokens to 25 million tokens. Under this free quota offer, the calling price for the personal version/entry-level GLM-3 Turbo model is also calculated in “cents.

SEE ALSO: Tongyi Qianwen’s Flagship Model Qwen-Long Announces A 97% Price Reduction

A Baidu insider said that recently, domestic and foreign competitors have successively released large-scale application products. DouBao also follows the route of public cloud + closed-source technology, consistent with Baidu. This indicates that closed-source large models are gradually becoming the mainstream in the market, which is the general trend of the industry. ‘At this stage, compared to price and performance, it’s much better to focus on open source rather than private ownership. What customers truly care about is effectiveness. In the era of generative AI, we should not only consider prices but also pay attention to application effects and response speed.’

On May 16, Baidu (Nasdaq: BIDU, 09888.HK) released its financial report for the first quarter of 2024, with a total revenue of 31.5 billion yuan, a year-on-year increase of 1%; Baidu‘s net profit attributable to shareholders (Non-GAAP) was 7.011 billion yuan, an increase of 22% year-on-year.

According to the information from the financial report conference call, Baidu AI Cloud achieved a revenue of 4.7 billion yuan in this quarter and continued to achieve profitability under non-US GAAP. At the same time, the proportion of revenue contributed by generative AI reached 6.9%, gradually expanding its share. There are two main factors behind the revenue growth: one is incremental income directly generated by generative artificial intelligence and basic models; second is new opportunities brought by traditional cloud business.

During the phone call, Baidu‘s founder, chairman and CEO Robin Li stated that the company is transitioning from being internet-centric to prioritizing artificial intelligence. They are advancing the reconstruction of To C and To B businesses using the ERNIE large model. Currently, the ERNIE large model processes approximately 250 billion tokens of text daily with an average of 200 million calls per day. Robin Li believes that search engines are most likely to become a “killer” application in the AI era, and Baidu‘s AI reconstruction work for search is still in its early stages.

Data shows that in terms of Baidu‘s intelligent cloud business, thanks to mutual promotion between ERNIE large models and PaddlePaddle, as of mid-April, there were already 13 million developers on the PaddlePaddle community platform. Industry insiders believe that generative artificial intelligence and basic model businesses are still at a very early stage. In the long run, Baidu‘s main focus remains on disseminating relevant knowledge and improving product adoption rates.