Baidu launches Ernie 5.1, a distilled sub‑model aimed at cutting pre‑training costs

Baidu has released Ernie 5.1, a substantially smaller sub‑model distilled from Ernie 5.0 and available through the company’s cloud and creative applications. The firm describes Ernie 5.1 as roughly one‑third the total parameter count of its predecessor and using about half the active parameters per query, a size and efficiency profile it says should lower operational cost for hosted deployments. For builders, that means a practical path to run multiple model sizes with less pre‑training overhead, although hosted access rather than downloadable weights constrains independent benchmarking and local use.

On public leaderboards and in company comparisons, Baidu reports several competitive results. As of May 9, Ernie 5.1 scored 1,223 points on the Search Arena leaderboard, placing fourth globally and first among Chinese models; a pre‑release preview scored 1,476 on the Text Arena, placing 13th. Baidu also reports Ernie 5.1 outperformed DeepSeek V4‑Pro on select autonomous agent benchmarks (tau3‑bench, SpreadsheetBench‑Verified) and nearly matched Gemini 3.1 Pro on GPQA and MMLU‑Pro in internal tests.

Baidu highlights a major compute‑efficiency gain in pre‑training Ernie 5.1: the company says the distilled model required only about 6% of the compute typically needed to pre‑train comparable standalone models. Baidu attributes that reduction to extracting a smaller sub‑model from a family trained once for multiple sizes, thereby reusing the heavy pre‑training investment made for Ernie 5.0 instead of running separate large pre‑training jobs for each configuration.

To support large‑scale fine‑tuning and reinforcement learning, Baidu rebuilt its RL infrastructure to run model updates, response generation and evaluation as decoupled subsystems coordinated by a central controller. Each component is matched to appropriate hardware to avoid pipeline stalls; Baidu also uses a standardized low‑precision computation library and a mixture‑of‑experts correction mechanism it says halves drift between training and example generation without noticeably slowing throughput.

Ernie 5.1 illustrates a route for organizations that want multiple maintained model sizes and isolated skill blocks while sharply cutting pre‑training cost, but it also exemplifies trade‑offs of hosted, closed‑weight releases. Baidu’s internal evaluations claim parity with Western commercial models on creative writing and competitive results on agentic and reasoning tasks, yet those claims remain tied to company‑run tests and access to Baidu’s hosted service.