Chinese artificial intelligence models continue to deliver impressive performance in terms of efficiency, competing head-to-head with their Western counterparts. The latest example of this is Baidu's new Ernie 5.1 model. According to data shared by the company, Ernie 5.1 offers high-level performance despite having a much lower training cost compared to similarly scaled models.
Ernie 5.1 features a significantly reduced structure compared to its predecessor, Ernie 5.0. The company states that the total number of parameters has been reduced by approximately one-third, and the number of effective parameters has been halved. Despite this, there has been no significant drop in the model's performance; in fact, it is said to have reached a level where it can directly compete with rival models in some areas. Baidu's most striking claim concerns training costs. According to the company, Ernie 5.1 was developed at only approximately 6% of the pre-training cost of similar models in the industry.
Baidu Adopts a Different Approach to Training AI
At the core of this efficiency increase is Baidu's method, which it describes as "multi-dimensional flexible pre-training." This approach, first used in Ernie 5.0, allows for the generation of sub-models of different scales within a single training process. For Ernie 5.1, Baidu reportedly built the new model by selecting the most efficient sub-network from the flexible sub-model pool created by Ernie 5.0. This allows the company to significantly reduce computational costs by reusing the most successful parts of the existing structure instead of training the model entirely from scratch.
Baidu also states that it used a new fine-tuning method called "asynchronously operating new reinforcement learning system" to develop the model. According to the company's statement, this system was particularly used to improve the model's performance in areas such as reasoning, tool usage, search-retrieval mechanisms, and multi-source content synthesis. This indicates that "AI agent" features, which have been gaining prominence recently, are also within the scope of Ernie 5.1. While Ernie 5.1 may not be a full-fledged AI agent, it can be seen as an indicator of the company's steps in this direction.
Ernie-5.1's Performance Test Results Are Impressive
The benchmark (performance test) results shared by Baidu indicate that the company is positioning Ernie-5.1 to directly compete with high-end models. In the LMArena Search Ranking, the model scored 1223 points, placing it fourth globally and first among Chinese models. The company also states that Ernie 5.1 surpassed DeepSeek-V4-Pro in agent evaluation tests and came very close to leading closed-source models in knowledge-focused benchmarks like GPQA and MMLU-Pro. Of course, these figures shared by Baidu need to be verified by independent sources, but the initial data is quite impressive.
Ernie 5.1 is currently available to developers and enterprise customers through Baidu's Qianfan platform and the Ernie website. The model is expected to become even more accessible in the coming period.
0 Comments: