Baidu Launches Ernie 5.1, Claims Top AI Performance at Just 6% Training Cost
A new AI foundation model has launched with a strong focus on efficiency rather than brute-force scaling. The latest release reportedly delivers near-frontier performance while using only a small fraction of the training cost typically associated with similarly capable models. As a result, this development highlights a growing shift toward smarter optimization in AI training.
The model significantly reduces its size compared to its predecessor. Specifically, total parameters have dropped to roughly one-third, while activated parameters have been cut in half. Despite this reduction, benchmark performance remains highly competitive across multiple domains.
On the LMArena Search Leaderboard, the model scored 1,223 on May 9, securing fourth place globally and leading among Chinese AI models. Meanwhile, on the Text Leaderboard, it ranked 13th worldwide with 1,476 points. Additionally, it placed within the top 10 in legal, government, mathematics, and business management benchmarks.
The company also reports strong results against competing systems. According to internal evaluations, the model outperformed DeepSeek-V4-Pro in agent-based benchmark tests, including τ³-bench and SpreadsheetBench-Verified. Furthermore, it approached top closed-source competitors in GPQA and MMLU-Pro evaluations. With tool use enabled, it scored 99.6 on AIME26, placing just behind Gemini 3.1 Pro. Its creative writing performance reportedly matches Gemini 3.1 Pro as well.
Training Strategy Behind the Gains
Rather than building the model entirely from scratch, developers adopted a more efficient strategy. First, they extracted an optimized sub-network from the earlier generation’s elastic sub-model matrix. Consequently, the new model inherited existing knowledge while dramatically reducing computational requirements.
Developers then applied additional optimization methods to strengthen performance. For example, they used decoupled fully-asynchronous reinforcement learning to refine reasoning capabilities. Moreover, they expanded agentic post-training to improve search retrieval, reasoning accuracy, and synthesis across multiple content sources.
This two-stage process appears central to the model’s cost advantage. Instead of relying solely on larger parameter counts, the strategy emphasizes targeted efficiency and smarter resource allocation.
Market Rollout and Industry Shift
The model is now available for enterprise users and developers through the company’s AI platforms. Additionally, more product announcements are expected at an upcoming developer conference next week.
This launch arrives as AI developers increasingly prioritize efficiency over raw scale. While Western competitors often depend on massive compute investments, alternative approaches now aim to achieve similar performance with far fewer resources. Therefore, this release may reflect a broader industry move toward cost-efficient AI development without sacrificing competitiveness.
