Baidu releases ERNIE 5.1

Background

Before ERNIE 5.1, Baidu's flagship model was ERNIE 5.0. The Decoder reported that ERNIE 5.0 has roughly 2.4 trillion total parameters, with fewer than three percent of them active for any single query^[3]. At that scale, pre-training is the single largest cost of a model generation, and how much a company spends on it shapes how often it can ship a new version.

The first public signal of the next version came from a leaderboard rather than a launch event. On 30 April 2026, a preview build called ERNIE-5.1-Preview ranked as the top Chinese model on the LMArena text leaderboard^[1][2]. LMArena is a public model comparison site where users vote on pairs of answers from different models, and the text leaderboard turns those votes into a ranking.

The release

Baidu officially released ERNIE 5.1 on 9 May 2026^[1][2]. The company did not train a new base model from scratch: ERNIE 5.1 was derived from ERNIE 5.0's pre-training^[3]. The result is considerably smaller, with about one-third of ERNIE 5.0's total parameters and half of its active parameters^[3]. Baidu said its pre-training cost was about 6 percent of that of comparable models^[3].

Total parameters and active parameters measure different things. Total parameters are the weights a model stores; active parameters are the ones that take part in computing a single answer. When a model's active count is far below its total, each query is routed through only part of the network, so the compute spent on one answer is smaller than the raw parameter count would suggest.

Scores and comparisons

On LMArena's Search Arena, ERNIE 5.1 scored 1,223 as of 9 May, placing 4th globally and 1st among Chinese models^[3][4]. Baidu also said that ERNIE 5.1 beats DeepSeek-V4-Pro on agentic tasks such as tau3-bench and SpreadsheetBench-Verified^[3]. This kind of agentic benchmark tests whether a model can carry out a multi-step task rather than simply answer a single question.

Hardware and showcase

Baidu showcased ERNIE 5.1 at Create 2026 on 13 May 2026, and said that a Kunlunxin chip cluster supported its training^[5]. That is, the chips behind the training run were named alongside the release itself, at an event held a few days after the model went out.

Why it matters

Baidu's first-quarter 2026 results release said that ERNIE 5.1 ranked first among Chinese models on LMArena's text leaderboard^[4]. Naming a specific model version in an earnings announcement is a way of putting a technical result into a business statement.

The more concrete claim is about cost. Cutting the pre-training bill to roughly 6 percent of comparable models, while shrinking the model to about a third of ERNIE 5.0's total parameters, is Baidu's answer to the question of how to keep releasing new versions without retraining a frontier-scale base model each time^[3]. ERNIE 5.1 inherits ERNIE 5.0's pre-training instead of replacing it, and the version number follows the earlier model rather than marking a fresh start^[3].

The release also follows a pattern that became visible in the weeks around it: the preview build appeared on a public leaderboard first, the ranking was cited in quarterly results, and the official launch came afterwards^[1][2][4]. For readers tracking how frontier models are compared, that sequence makes public arenas part of the announcement itself, alongside the vendor's own benchmark comparisons against DeepSeek-V4-Pro^[3].