ERNIE 5.1

ERNIE 5.1, announced by Baidu on 8 May 2026 and rolled out more widely at the Create 2026 developer conference in Beijing the following week, is the model with which Baidu stopped competing on Chinese-language tasks alone and started competing on cost. The headline number is not a benchmark but a training bill. Baidu says ERNIE 5.1 reaches its level of foundational performance using roughly 6% of the pre-training compute of comparable models, and that it did so while shrinking the model: total parameters fall to about one third of ERNIE 5.0 and active parameters per token to about one half. The technique Baidu credits is multi-dimensional elastic pre-training, in which depth, expert capacity and sparsity are varied during training rather than fixed in advance — a sparse mixture-of-experts model that is, in effect, allowed to change shape while it learns. The results held up on public leaderboards. On the Arena Search leaderboard ERNIE 5.1 scored 1,223, fourth in the world and first among Chinese models. It scores 99.6 on the AIME 2026 mathematics set, behind only Gemini 3.1 Pro, and Baidu reports it ahead of DeepSeek-V4-Pro on the agentic tool-use benchmark tau-3-bench and on SpreadsheetBench, with GPQA and MMLU-Pro results it describes as approaching leading closed-source models. What that adds up to is a shift in what Chinese frontier models are arguing about. ERNIE 5.0, launched in November 2025 alongside Baidu's claim of 200 million monthly users of its assistant, was a scale story. ERNIE 5.1 is an efficiency story, and it lands in the same year that DeepSeek, Alibaba and Moonshot are all making the same argument in different words: that the frontier can be reached without frontier spending. The model is closed-weight. It is reachable through the ERNIE Bot web interface, an ERNIE 5.1 playground in Baidu AI Studio, and the Qianfan API, and Baidu lists deployment across more than ten third-party creative platforms.

▸Videos

No videos yet.

▸Reviews