Baidu's Ernie 5.1 cuts 94 percent of pre-training costs while competing with top models

The 94 percent figure is relative to Baidu's own prior training runs, not an independently audited industry baseline, which makes the headline number harder to benchmark against what Google, Anthropic, or OpenAI actually spend per training run. Fourth place on Search Arena is competitive, but Search Arena skews toward retrieval-augmented tasks where Baidu has structural advantages through its search index.

Modelwire has no prior coverage to anchor this to directly, so the honest framing is that this belongs to a broader pattern worth tracking: Chinese labs finding efficiency routes around the compute constraints imposed by US export controls on high-end chips. Baidu cannot freely access H100-class hardware at scale, which creates a genuine incentive to extract more from fewer FLOPs. The 'Once-For-All' approach, producing multiple sub-models from one training run, reads less like a philosophical commitment to efficiency and more like an adaptation to a constrained supply environment.

If Ernie 5.1's sub-model variants hold their Search Arena rankings on third-party evaluations outside Baidu's own reporting pipeline within the next two quarters, the efficiency claim becomes credible. If independent evals show significant degradation, the 94 percent cost reduction likely came with quality trade-offs the current benchmarks don't surface.