Baidu releases smaller, cheaper ERNIE 5.1
Baidu's ERNIE 5.1 has about a third of ERNIE 5.0's total parameters, and the company says pre-training cost about 6% of what comparable models need while performance stays near flagship level.
Baidu announced the release of ERNIE 5.1, a model built on ERNIE 5.0's pre-training rather than trained from scratch. ERNIE 5.1 has roughly one-third of its predecessor's total parameters and about half of its active parameters. Baidu says pre-training compute came to about 6% of what comparable models at the same scale require.
The savings come from what Baidu calls a "Once-For-All" elastic training framework. ERNIE 5.0 was trained once while the system sampled sub-models of different depths, expert capacities and routing sparsity (via variable Top-k expert activation). ERNIE 5.1 is the best sub-network taken from that set. Baidu says this cuts inference cost compared with ERNIE 5.0.
For post-training, Baidu built a new reinforcement learning system on PaddlePaddle that runs fully asynchronously. Its training, inference, reward and agent-loop subsystems are split apart and can each be deployed and scaled independently.
Baidu's benchmark claims:
- Search Arena: a score of 1,223 on May 9, placing 4th overall and 1st among Chinese models.
- Agent tasks: ahead of DeepSeek-V4-Pro on τ³-bench and SpreadsheetBench-Verified.
- Math: 99.6 on AIME26 with tool use, second only to Gemini 3.1 Pro.
- Knowledge and writing: close to leading closed-source models on GPQA and MMLU-Pro. Baidu's internal tests put its creative writing near Gemini 3.1 Pro.
Most of these figures are Baidu's own, and the post gives no exact parameter counts. ERNIE 5.1 is available to chat with at ernie.baidu.com and in a Playground on Baidu AI Studio. The announcement says nothing about open weights or API pricing.