Baidu shrinks Ernie 5.1 by 66%

Baidu’s ERNIE 5.1 story is really about a shift in how AI models get built. For the last two years, the obvious move was simple — make the model bigger, spend more, and hope capability rises with scale. But that path is getting brutally expensive. Baidu is now arguing that the smarter move, at least for production, is to keep most of the capability while cutting a huge amount of the model and training bill. ERNIE 5.1, released on May 9, is the clearest version of that pitch yet. ### What actually changed in ERNIE 5.1? Baidu says ERNIE 5.1 inherits the pretraining base of ERNIE 5.0, then compresses the system so total parameters fall to roughly one-third of the earlier model and active parameters fall to roughly one-half. In plain English, the model is not starting over from scratch with a giant new brain. It is taking the previous foundation, keeping the useful structure, and trimming the rest hard enough to change the economics. (ernie.baidu.com) ### Why is “active parameters” the important part? A model’s total parameters tell you how large it is overall. Active parameters tell you how much of that model gets used when it answers a prompt. That second number matters more for serving cost and latency. If Baidu really cut active parameters by about 50%, the win is not just cheaper training once — it can also mean cheaper real-world usage every day for search, agents, and enterprise apps. That is the production angle engineers care about. (ernie.baidu.com) ### Where does the 94% cheaper claim come from? It comes from Baidu’s statement that ERNIE 5.1 uses only about 6% of the pretraining cost of comparable models at similar scale. Flip that around and you get a roughly 94% reduction. The catch is that this is Baidu’s own comparison set, not a neutral industry audit. Still, even as a directional claim, it is a big one. It says the company thinks model design and training strategy now matter as much as raw compute spend. (ernie.baidu.com) ### How is Baidu saying it pulled that off? The company points to two main ideas. First, “multi-dimensional elastic pre-training,” which basically means extracting efficient sub-structures from ERNIE 5.0 rather than scaling a fresh giant model. Second, a new fully asynchronous reinforcement learning setup for post-training, aimed at reducing wasted compute and handling messy long-tail tasks better. The jargon is dense, but the idea is familiar — spend less time training parts of the system that do not buy much extra intelligence. (ernie.baidu.com) ### Did capability actually hold up? Baidu says yes, and the headline benchmark it is pushing is search. ERNIE 5.1 scored 1,223 on Arena Search on May 9, which Baidu says put it fourth globally and first among Chinese models. Baidu also says the model beats DeepSeek-V4-Pro on some agent evaluations and comes close to leading closed models on knowledge, reasoning, and writing tasks, including Gemini 3.1 Pro on some comparisons. Those are strong claims, but they are still mostly benchmark claims — not the same thing as broad real-world superiority. (ernie.baidu.com) ### Why does this matter beyond Baidu? Because this is where the market is heading. Frontier labs still want giant models for prestige and absolute performance. But companies shipping AI products want something else — lower inference cost, lower training cost, and fewer GPUs tied up per useful task. ERNIE 5.1 fits that second path. It is less a “we built the biggest thing” announcement and more a “we found a better cost-performance point” announcement. (ernie.baidu.com) ### So is this a real breakthrough? Probably yes on efficiency, with one caveat. The numbers are coming from Baidu’s own release, so they need outside validation over time. But the broader signal is real either way. AI development is no longer just a race to add parameters. It is becoming a race to decide which parameters are actually worth paying for. (ernie.baidu.com)