--- title: "Baidu's new model does more on a fraction of the training bill" date: 2026-10-22 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-ernie-51 language: en --- # Baidu's new model does more on a fraction of the training bill > Baidu's ERNIE 5.1 keeps flagship-level skill while cutting model size and training cost sharply, and is tuned to act as an agent. The bill for training a top-tier model has become the quiet bottleneck of the whole industry. Teams that want a model competitive with the best must spend on scarce compute (算力), and the cost compounds with every experimental run. Baidu says it found a cheaper road this spring. Its newest release keeps much of the skill of its previous giant while shrinking the model and the training budget at the same time. The system is called ERNIE 5.1, and it shows how a Chinese large model (大模型) maker is betting on efficiency as much as raw size. ## The efficiency claim Baidu released ERNIE 5.1 in May 2026. Compared with its earlier ERNIE 5.0, the company says it compressed total parameters to about one-third and active parameters to about one-half, while still matching strong models on many tasks. The sharper claim is about cost. Baidu states that the pre-training compute needed for ERNIE 5.1 was only about 6% of what comparable models required. In other words, the same level of capability, reached with a much smaller bill. Why this lands with enterprises is straightforward. A model that costs less to train can also be retrained often, which means a company can keep a private version current without a hyperscaler's budget. Those are Baidu's own figures, and they describe the company's internal accounting rather than an outside audit. But the direction is the interesting part: capability per dollar, not parameters per se. ## How Baidu says it got there The technique has a name inside the company: an "Once-For-All" elastic training framework. Instead of training one model at one size, Baidu says it trained a large family of sub-models together in a single run, then pulled out the best-sized network for ERNIE 5.1. That lets the team inherit the knowledge baked into the bigger ERNIE 5.0 while avoiding a full, expensive pre-training from scratch. The model also uses a mixture-of-experts structure, where only a slice of the network activates per request, which keeps serving cheap. The payoff is not only cost. A model that can be shrunk and grown along several dimensions gives Baidu a single training run that yields many products — a lightweight version for phones, a heavier one for the cloud — instead of separate efforts. ## Tuned to act, not just answer Like its peers, Baidu is pushing the agent (智能体) angle. It built a new reinforcement-learning setup described as fully asynchronous, meant to handle the gap between training and inference and to stabilize long, messy tasks. On agent (智能体) evaluations, Baidu reports that ERNIE 5.1 beats DeepSeek-V4-Pro on τ³-bench and SpreadsheetBench-Verified, and that its maths score on a competition benchmark reached 99.6 with tool use. On a search-oriented leaderboard, it says the model scored 1,223 to rank fourth globally and first among Chinese models. Again, these are Baidu's own reports. They tell you where the team is aiming, not a final verdict. ## Why efficiency is the real story A model that costs a fraction to train changes who can play. A smaller player with a smart training method can field a capable model without a hyperscaler's budget. It also changes how often a lab can iterate — cheap experiments mean faster learning. China's model makers have spent 2026 competing as much on price as on leaderboards, and Baidu's efficiency story fits that wider trend. A cheaper model is also a more deployable one. For users, a cheaper-to-serve model usually means lower prices at the interface, which is why efficiency claims from Chinese labs matter even if you never call ERNIE directly. ## The competitive read ERNIE 5.1 enters a Chinese market where several labs now ship models that claim both high quality and low cost. Baidu's bet is that efficiency, not just leaderboard position, wins enterprise adoption, because a cheaper model is far easier to embed in products at scale. The claim also reframes the race: instead of asking which model is biggest, the useful question becomes which model delivers enough skill per unit of compute (算力). ## Honest limitations The 6% training-cost figure and the benchmark scores are Baidu's own and have not been independently verified. The comparison to DeepSeek-V4-Pro is framed by Baidu and may not hold on every task. Efficiency at training time does not guarantee low latency or low cost at serving time in practice. Independent English-reasoning tests for ERNIE 5.1 are still limited. ## What readers can do now - Read Baidu's technical post on the elastic training method if you care about cutting pre-training cost; the idea transfers beyond one model. - Test ERNIE 5.1 on a real spreadsheet or search task through Baidu's platform before trusting the agent (智能体) claims. - Compare per-token prices across Chinese and non-Chinese hosts rather than assuming bigger models cost more to use. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-ernie-51 Free to quote with attribution and a link to the original.