Baidu's New AI Is Already Beating Top Models and Cost 94% Less to Build

6 days ago 13

In brief

  • Baidu says ERNIE 5.1 pre-training outgo is conscionable 6% of what comparable AI models spend
  • Despite this, the exemplary is ranked 4th globally connected the LMArena Search leaderboard.
  • Baidu compressed the exemplary to astir one-third the full parameters of its predecessor, ERNIE 5.0, without sacrificing flagship-level performance.

Baidu, China's equivalent of Google, officially released ERNIE 5.1 precocious past week, and the header fig is hard to ignore: The exemplary outgo astir 94% little to bid than comparable AI systems astatine the aforesaid scale.

To enactment that successful perspective, grooming a frontier AI exemplary typically costs respective millions (if not billions) of dollars successful compute. Baidu, which controls implicit 76% of China's hunt marketplace and trades connected Nasdaq arsenic BIDU, is claiming it pulled disconnected the aforesaid people of show for astir a twentieth of that.

The instrumentality is called "multi-dimensional elastic pre-training." Instead of gathering ERNIE 5.1 from scratch, Baidu extracted an optimized sub-network from its existing ERNIE 5.0 architecture—which it released successful January 2026—and compressed it down. Total parameters dropped to astir one-third of the original. Active parameters (the ones really doing enactment during a conversation) were chopped successful half. The effect is simply a leaner exemplary that inherited the cognition basal of its larger genitor without repeating the afloat grooming bill.

On LMArena Search Arena—a leaderboard wherever existent users comparison AI models connected unrecorded web hunt tasks, scored by quality preference—ERNIE 5.1 scored 1,223, landing 4th globally and archetypal among each Chinese models. Its agentic capabilities—how good it handles multi-step tasks similar filling spreadsheets oregon autonomously browsing the web—surpassed DeepSeek-V4-Pro, the erstwhile Chinese benchmark-setter.

ERNIE whitethorn dependable similar an exotic sanction among occidental observers, but it’s really a large exemplary successful China. Baidu launched Ernie Bot successful August 2023, and the chatbot reached 100 cardinal users successful China by December of that year—faster than astir Chinese rivals, though inactive slower than ChatGPT's planetary grounds two-month tally to the aforesaid milestone.

The ratio space echoes what DeepSeek did to the AI manufacture successful January 2025. When the Chinese startup released R1—a exemplary that matched OpenAI's o1 astatine 98% little query cost—it triggered a $600 cardinal wipeout successful Nvidia's marketplace worth and forced each large AI laboratory to rethink whether throwing compute astatine problems was the lone viable strategy. ERNIE 5.1 is simply a antithetic benignant of ratio story—on the grooming broadside alternatively than inference—but the underlying connection is the same: Chinese labs support uncovering ways to bash much with less.

The post-training pipeline is besides worthy noting. Baidu built a four-stage reinforcement learning strategy it calls MOPD (Multi-Teacher On-Policy Distillation). Rather than trying to thatch each accomplishment astatine once—which tends to origin "seesaw effects" where, for example, improving mathematics show tanks originative writing—Baidu trained specializer adept models successful parallel for code, reasoning, and agentic tasks, past distilled each of them into a azygous unified model. A last online reinforcement learning signifier handled open-ended conversations and originative output, preserving what the distillation process couldn't seizure well.

In mentation it should mean each skills are leveled successful presumption of proficiency, without 1 being prioritized implicit the other

On GPQA (Graduate-Level Google-Proof Q&A, a benchmark measuring whether a exemplary tin reply expert-level subject questions that can't beryllium Googled), ERNIE 5.1 approaches the show of starring occidental closed-source models. On AIME26—the American Invitational Mathematics Examination adapted for 2026, which tests precocious problem-solving nether contention conditions—the exemplary scored 99.6% erstwhile utilizing tool-assisted reasoning, trailing lone Gemini 3.1 Pro.

Baidu says ERNIE 5.1 is already rolling retired crossed much than 10 originative and agentic platforms successful China, including AI roleplay platforms and abbreviated play procreation tools. The exemplary is accessible astatine ernie.baidu.com and via API connected Baidu's AI Cloud platform.

Baidu is hosting its yearly Create 2026 developer league connected May 13–14 successful Beijing, wherever it plans to showcase ERNIE's latest concern applications. That lawsuit volition beryllium the adjacent information constituent connected however aggressively the institution intends to propulsion the exemplary into endeavor and planetary markets.

Daily Debrief Newsletter

Start each time with the apical quality stories close now, positive archetypal features, a podcast, videos and more.

Read Entire Article