ERNIE 5.1

What it is

ERNIE 5.1 is a large language model released by Baidu on 9 May 2026^[1]. Baidu frames the release around text: it describes stronger text capabilities, a more compact size and enhanced reasoning^[4]. It is a hosted service rather than an open model — it can be used through ernie.baidu.com and a playground in Baidu AI Studio, and no model weights were released^[3].

Where it came from

The model it builds on, ERNIE 5.0, is very large: roughly 2.4 trillion total parameters, fewer than three percent of which activate per query^[3]. Parameters are the numbers inside a model that training adjusts; the active ones are the subset that actually takes part in computing a single answer, while the rest sit idle for that request. ERNIE 5.1 keeps ERNIE 5.0's pre-training foundation but is far smaller, with about one-third of the total parameters and about half the active parameters per query^[3].

Baidu said the pre-training cost was about 6% of what comparable models require. The method it names is "Once-For-All", which derives smaller models from ERNIE 5.0's compute instead of paying for a separate, from-scratch training run^[3]. Pre-training is normally the most expensive stage of building a language model, so reusing an existing run changes the economics of shipping a mid-sized model: capability is inherited from a larger predecessor rather than bought again.

The public sequence ran in two steps. ERNIE-5.1-Preview ranked as the No. 1 Chinese model on the LMArena text leaderboard on 30 April 2026, and the official release followed on 9 May 2026^[1][2].

What it does, and how it was measured

On LMArena, ERNIE 5.1 ranked first among Chinese models on both the text and search leaderboards, and fourth globally on search^[4]. On the Search Arena it scored 1,223 points as of 9 May 2026^[3]. Baidu's charts showed 99.6 on AIME26 with tools and 72.5 on SpreadsheetBench-Verified^[3].

Baidu also says ERNIE 5.1 beats DeepSeek-V4-Pro on tau3-bench and SpreadsheetBench-Verified, and comes close to Gemini 3.1 Pro on GPQA and MMLU-Pro^[3]. These are vendor comparisons: leaderboard placements come from a third-party platform, while the head-to-head claims are Baidu's own, and both kinds of number move as models and test sets are updated.

The "with tools" qualifier matters for reading the AIME26 figure. A score measured with tool use describes a system that can call external computation during a problem, not a model answering entirely on its own, so it is not directly comparable to a no-tools number.

Compute behind it

A dedicated Kunlunxin chip cluster supported training of a key model in the ERNIE 5.1 series^[5]. Naming the hardware alongside the model points at the full stack — chips, cluster, training run — behind a release, not just the finished software.

Where it sits

ERNIE 5.1 is a small-model-from-big-model release: capability inherited from a much larger predecessor, cost cut by derivation rather than by a fresh training campaign^[3]. It is also closed — usable as a service, not downloadable^[3]. And it arrives into a crowded comparison set, where Baidu measures it against DeepSeek-V4-Pro and Gemini 3.1 Pro, and where leaderboard placements are read as a rough signal of standing among Chinese models^[3][4]. What it shows is a direction of travel as much as a single model: less spent on pre-training, more on reusing what a larger run already produced.