Baidu on Friday released Ernie 5.1, its next-generation foundation model, claiming near-frontier performance at a fraction of the computational expense typically required to train models of comparable scale. The release marks a continued push by Chinese AI developers to close the gap with Western competitors through efficiency rather than brute-force scaling.
Smaller, Cheaper, Competitive
Ernie 5.1 compresses total parameters to roughly one-third and activated parameters to one-half of its predecessor, Ernie 5.0, which contained 2.4 trillion parameters. Through what Baidu calls “multi-dimensional elastic pretraining” — a technique first introduced with Ernie 5.0 that enables a single training run to produce models at multiple scales — the company says it achieved these results using only about 6% of the pretraining cost of comparable industry models.
On the LMArena Search Leaderboard, the model scored 1,223 on May 9, placing fourth globally and first among Chinese models. On the Text Leaderboard, it ranked 13th globally with 1,476 points, also placing in the top 10 across legal, government, math, and business management categories. Baidu says Ernie 5.1 surpasses DeepSeek-V4-Pro on agent evaluation tasks including τ³-bench and SpreadsheetBench-Verified, approaches leading closed-source models on the GPQA and MMLU-Pro knowledge benchmarks, and scores 99.6 on AIME26 with tool use — second only to Gemini 3.1 Pro. Its creative writing capabilities, according to Baidu’s internal evaluations, match those of Gemini 3.1 Pro.
Technical Approach
The efficiency gains stem from a two-stage process. Rather than training from scratch, Ernie 5.1 extracts an optimal sub-network from Ernie 5.0’s elastic sub-model matrix, inheriting its predecessor’s knowledge while dramatically cutting compute. Baidu further refined the model using “decoupled fully-asynchronous reinforcement learning” and scaled agentic post-training to improve reasoning, search retrieval, and multi-source content synthesis.
Availability and Competitive Context
Ernie 5.1 is now available to enterprise users and developers through Baidu’s Qianfan model platform and the Ernie website. Baidu’s official Facebook page noted that additional product launches are expected at its Create 2026 conference next week.
The release comes as efficiency-focused development gains momentum across the Chinese AI industry, with companies seeking to deliver competitive performance without the massive compute budgets of Western rivals. Ernie 5.1’s cost profile — achieving top-tier rankings while spending a small fraction of what comparable models require — represents what one analysis described as “a shift toward efficiency-driven post-training optimization over pure parameter scaling”.
