ERNIE 5 by Baidu: status and timeline

ERNIE 5

Give China a homegrown frontier model that matches the best US systems, and turn Baidu's search audience into users of AI agents that do tasks for them.

ScalingStage 5 of 5

ERNIE 5.1, released May 2026, is Baidu's flagship; it placed 4th on LMArena's search leaderboard at launch.

Updated May 2026Checked 11 Oct0 updates this week

Milestones

No announced next step

ERNIE Bot opens to the public in ChinaAug 2023Complete.
ERNIE 4.5 family released as open weights30 Jun 2025Complete.
ERNIE 5.0 preview unveiled at Baidu World13 Nov 2025Complete.
ERNIE 5.0 enters LMArena text top 10 (#8)15 Jan 2026Complete.
ERNIE 5.0 official releaseJan 2026Complete.
ERNIE 5.1 releasedMay 2026Complete.

Most important updates

Current obstacles

No outside verificationWeights stay closed, so scores and the 94% training-cost saving rest on Baidu's word and leaderboard votes.
Behind on general text rankingsERNIE 5.1 placed 13th in a preview of LMArena's text board, behind US leaders, despite its strong search ranking.

Physics limits

Memory, not maths, caps sparse modelsOnly a few experts run per answer, but all 2.4 trillion parameters must sit in fast chip memory, so serving still needs large, costly clusters.
Good text data is running outModels improve with more high-quality text, but the stock written by people is finite; labs turn to synthetic data, video and audio, which teach less per byte.
Chip export limits cap computeUS controls keep top Nvidia chips out of China, so Chinese labs train on fewer or domestic chips and must win on efficiency instead of scale.

How it works

3 parts

Experts

Many specialists, few used at once

The model is split into many expert sub-networks. A router sends each piece of input to a few of them, so under 3% of the 2.4T parameters work per answer.

Senses

Native text, image, audio, video

All four kinds of data are trained together from scratch, so one model can read a video and answer in speech without hand-offs between models.

Elastic

Train once, cut to size

ERNIE 5.1 was cut out of ERNIE 5.0 with a method that trains differently sized versions together; Baidu kept a smaller one that costs less to train and to run.

Spec sheet

Spec	ERNIE 5
Total parameters (5.0)	2.4 trillion^R (reported)
Share active per answer (5.0)	Under 3%^R (reported)
ERNIE 5.1 size vs 5.0	About 1/3 total, 1/2 active parameters^R (reported)
5.1 pre-training cost	About 6% of comparable models^R (reported)
LMArena text score (5.0, Jan 2026)	1,460 (#8)^R (reported)
Search Arena score (5.1)	1,223 (#4)^R (reported)
API calls per day (Jan 2026)	16.5 billion^R (reported)

^R reported by the company

Papers & demos

1 item

30 Jun 2025code
ERNIE 4.5 family released as open weights (Apache 2.0) (opens technode.com)Ten models up to 424B parameters; Baidu's turn from closed models after DeepSeek's rise.

Update log

5 updates

About Baidu

The team behind ERNIE 5

Baidu

ERNIE models, Ernie assistant, AI search

Baidu runs China's largest search engine and builds the ERNIE family of AI models, which power its Ernie assistant, its search results and its cloud. It also owns the Kunlunxin chip unit and the Apollo Go robotaxi service, tracked separately.

ERNIE 5.0 (2.4 trillion parameters) was the only Chinese model in LMArena's text top 10 in January 2026.
ERNIE 5.1 (May 2026) is about a third the size of 5.0; Baidu says pre-training cost about 6% of comparable models.
The Ernie assistant had 202 million monthly users in December 2025; Baidu said ERNIE served 16.5 billion API calls a day in January 2026.

Company profile →

Founded: 200026 yrs
Headquarters: China
Status: Public
Listed: BIDU (opens google.com)NASDAQ
Valuation: public
Staff: ~33,500Jan 2025
Works in: AIFrontier models
Coverage: 1 program · 5 updateslatest 1 May 2026checked 11 Oct
Partners: JD.com (opens jd.com)Meituan (opens meituan.com)Trip.com (opens trip.com)
Suppliers: Baidu Kunlunxin
People: Robin Li (opens en.wikipedia.org)Co-founder, Chairman and CEO

Spec

ERNIE 5

Total parameters (5.0)

2.4 trillion^R (reported)

Share active per answer (5.0)

Under 3%^R (reported)

ERNIE 5.1 size vs 5.0

About 1/3 total, 1/2 active parameters^R (reported)

5.1 pre-training cost

About 6% of comparable models^R (reported)

LMArena text score (5.0, Jan 2026)

1,460 (#8)^R (reported)

Search Arena score (5.1)

1,223 (#4)^R (reported)

API calls per day (Jan 2026)

16.5 billion^R (reported)

ERNIE 5

Milestones

Most important updates

Current obstacles

Physics limits

How it works

Many specialists, few used at once

Native text, image, audio, video

Train once, cut to size

Spec sheet

Papers & demos

Update log

May 2026

Tue 20 Jan

Thu 15 Jan

Thu 13 Nov 2025

Mon 30 Jun 2025

About Baidu

Current obstacles

Physics limits

How it works

Many specialists, few used at once

Native text, image, audio, video

Train once, cut to size

Spec sheet

Papers & demos

Update log

May 2026

Tue 20 Jan

Thu 15 Jan

Thu 13 Nov 2025

Mon 30 Jun 2025

About Baidu