Baidu ERNIE 5.0 Launch: China's Most Powerful AI Model Yet

Baidu's ERNIE 5.0 was previewed on November 13, 2025 and released officially in January 2026. It is a 2.4-trillion-parameter model trained on text, images, audio and video together. This explainer covers what is confirmed, what the benchmark claims really show, and how to test it for your own work.

What ERNIE 5.0 is

ERNIE 5.0 is Baidu's flagship foundation model. Baidu previewed it at its Baidu World conference on November 13, 2025, and released it officially in January 2026. According to The Decoder, it has roughly 2.4 trillion parameters in a mixture-of-experts (MoE) design, and fewer than 3% of them are active for any single query. That is how a model this large can still be served at a reasonable cost: each request only runs through a small slice of the network.

What "native multimodal" means here

Many multimodal models start as text models and have vision or audio encoders bolted on later. Baidu says ERNIE 5.0 is different. Its CTO Haifeng Wang described a "unified auto-regression architecture for native full multimodal modelling", with speech, images, video and audio in the training data from the start (InfoWorld). In practice, the claim is that the model reasons across a chart, a screenshot or a clip as naturally as across text. Baidu has not published a full technical report, so outsiders cannot yet check how the architecture actually works.

The benchmark claims, read carefully

• LMArena: the preview version ranked joint second on the text leaderboard at launch, then slipped to eighth (InfoWorld). The released ERNIE-5.0-0110 scored 1,460 in the January 15, 2026 snapshot: 8th overall, level with GPT-5.1 (High), and 1st among Chinese models (The Decoder).
• Math: 2nd worldwide in LMArena's math category at that snapshot, behind GPT-5.2 (High).
• Charts and documents: Baidu says the model beats GPT-5 on chart and document understanding (VentureBeat). These are Baidu's own numbers.

LMArena rankings come from crowd votes and move week to week, and Baidu's own benchmark tables have not been independently reproduced. One analyst quoted by InfoWorld said the claims still need validating on independent benchmarks. Treat ERNIE 5.0 as a serious frontier-class model, but don't treat any single leaderboard position as settled.

Where it fits in Baidu's stack

ERNIE 5.0 is proprietary and built on Baidu's PaddlePaddle framework. That is a change from its smaller predecessor, ERNIE-4.5-VL-28B-A3B-Thinking, which was released under the Apache license. At the same event, Baidu announced its own inference and training chips: the Kunlunxin M100, due in early 2026, and the M300, due in early 2027. It also announced "Tianchi" supernodes built from hundreds of Kunlunxin chips. Together they show a company trying to own the whole stack, from chips to model to apps, as access to US accelerators stays restricted.

Our take: how to evaluate ERNIE 5.0 for real work

1. Start with your own documents. Baidu's strongest claims are about charts, tables and documents. Run 20 to 30 of your real PDFs or screenshots through it, and through the model you use today.
2. Check Chinese and English separately. A model can rank highly overall and still behave differently in each language. Score both if you serve both.
3. Read the data terms. It is a hosted, closed model. Confirm where prompts are processed and kept before you send anything sensitive.
4. Don't buy on rank. An eighth place on LMArena is a snapshot. Re-test when Baidu ships the next version, rather than relying on old leaderboard screenshots.

For a comparison point on a Western frontier model, see our GPT-6 Astra explainer.

Sources

• InfoWorld: Baidu launches new generation of Ernie AI
• The Decoder: ERNIE 5.0 tops all Chinese models in LMArena
• VentureBeat: Baidu unveils proprietary ERNIE 5
• Baidu ERNIE (official site)

Frequently asked questions

Related articles

Comments