Baidu ERNIE 5.1: Why the 6% Training-Cost Claim Matters

Baidu’s ERNIE 5.1 is best read as an efficiency story, not a bigger-model story. In its release, Baidu says ERNIE 5.1 inherits the pre-training foundation of ERNIE 5.0, compresses total parameters to approximately one-third and active parameters to approximately one-half, and achieves leading foundational performance at its model scale using only about 6% of the pre-training cost of comparable models .

That makes the announcement strategically interesting: Baidu is presenting a path to strong model performance that depends less on a fresh giant pre-training run and more on reuse, compression, and post-training. The 6% figure is compelling, but it should be treated as a company-reported claim until the comparison baseline and accounting details are clearer .

What Baidu actually claimed

Baidu’s central claim is narrow but important. ERNIE 5.1 is described as inheriting ERNIE 5.0’s pre-training foundation rather than being trained as an entirely new foundation model from scratch . Baidu says that, while doing so, it reduced total parameters to about one-third and active parameters to about one-half .

The cost claim is also specifically about pre-training. Baidu says ERNIE 5.1 uses only about 6% of the pre-training cost of comparable models . The provided materials do not establish that this figure covers total development cost, post-training cost, deployment cost, inference cost, hardware efficiency, or commercial pricing.

Baidu’s broader blog also says ERNIE 5.1 delivers upgrades across agent, reasoning, and creative capabilities, powered by disaggregated fully-asynchronous reinforcement learning and scaled agentic post-training . The same blog says the model ranked first in China on Arena Search Arena .

Why the 6% number matters

The AI model race is often discussed in terms of scale: more parameters, more data, more compute. ERNIE 5.1 points in a different direction. Baidu is claiming that meaningful performance can be preserved while shrinking the model footprint and avoiding a full-cost pre-training cycle .

If that approach holds up in practice, the competitive advantage shifts toward cost-performance engineering: how well a lab can reuse a foundation, select efficient sub-models, compress active computation, and improve behavior through post-training. ERNIE 5.1 matters because Baidu is making that argument explicitly through its release materials .

How Baidu says ERNIE 5.1 became cheaper

Baidu’s efficiency claim rests on four related ideas.

1. Reuse the ERNIE 5.0 foundation

The release says ERNIE 5.1 inherits the pre-training foundation of ERNIE 5.0 . That is the core of the cost argument: ERNIE 5.1 is positioned as a model derived from an existing foundation rather than a standalone, full-price pre-training effort.

2. Shrink both total and active parameters

Baidu says ERNIE 5.1 compresses total parameters to approximately one-third and active parameters to approximately one-half . Total parameters describe the full model footprint, while active parameters are the portion used in a given computation. Reducing both is why the release is as much about efficiency as capability.

3. Use elastic model-family training

The ERNIE 5.0 technical report describes an “elastic training” approach in which a single pre-training run can produce a family of models with different capacity-efficiency trade-offs . The report says this is done by dynamically sampling sub-models with different depth, width, and routing sparsity, and by allowing sub-models to inherit knowledge from the full model for later post-training stages .

That matters for ERNIE 5.1 because it helps explain the model-family logic behind Baidu’s claim. The reported method is not simply “train a bigger model.” It is closer to training a flexible foundation and then deriving more efficient configurations from it .

4. Add agentic and reasoning post-training

Baidu says ERNIE 5.1 uses disaggregated fully-asynchronous reinforcement learning and scaled agentic post-training to improve agent, reasoning, and creative capabilities . In other words, Baidu’s claim is not only that it made the model smaller; it is also saying post-training contributed to the final capability profile .

What remains unproven

The biggest open question is verification. The cited public materials do not provide a full public accounting of the training budget, hardware setup, data mixture, training duration, accelerator utilization, post-training cost, or exact set of “comparable models” behind the 6% figure .

That does not make the claim meaningless. It does mean the number should not be treated as an independently audited industry benchmark. The strongest supported reading is narrower: Baidu says ERNIE 5.1 preserved leading foundational performance at its model scale while reducing parameters and pre-training cost through inheritance, compression, elastic training ideas, and post-training .

The bottom line

ERNIE 5.1 is significant because it reframes Baidu’s AI progress story around cost-performance rather than raw scale. Baidu says the model inherits ERNIE 5.0’s foundation, cuts total and active parameters, and reaches leading foundational performance at its scale with about 6% of the pre-training cost of comparable models .

The claim is important, but it is not fully settled by the public materials. Until Baidu or independent evaluators disclose more about the baseline, hardware, data, and accounting behind the 6% figure, ERNIE 5.1 should be viewed as a serious efficiency claim—not a completely verified cost benchmark.