Every ERNIE AI Model Explained and Compared (Nov, 2026)

The Current ERNIE Models

Baidu's naming used to split into a general line (ERNIE 4.5, 4.5 Turbo) and a reasoning line (ERNIE X1, X1 Turbo, X1.1). That split is gone.

ERNIE 5.0 merged text, image, audio and video into one model, ERNIE 5.1 replaced X1.1 on Qianfan, and the newest release, 5.5 Flash, thinks on every call.

ERNIE-Light-DT-1.0 is the odd one out: a long-context model that Baidu's OpenRouter listing describes as built on open-source GLM models, the line our GLM model guide covers.

ERNIE 5.5 Flash Preview

Baidu's release page describes an 11-hour run in which the model wrote an Unreal Engine 5.8 wuxia game in C++, plus nearly 100 mobile mini-games built from design documents.

ERNIE 5.5 Flash Preview launched on November 11, 2026 as the first post-trained model on top of the 5.5 Flash Base checkpoint.

Baidu calls it an intermediate checkpoint and says some capabilities "may still fall short" of the official release.

It is the cheapest ERNIE on Qianfan: CNY 1 input and CNY 4 output per million tokens from 08:00 to 22:00, and half that overnight, both at a promotional rate against a list price twice as high. Cache hits cost a tenth of the input rate.

On OpenRouter the preview is free, text in and text out only. Baidu has published no benchmark scores for the preview in text form; its release page shows them only as images.

ERNIE 5.1

ERNIE 5.1 has about a third of ERNIE 5.0's total parameters and half its active ones, and Baidu says its pre-training cost about 6% of comparable models at the same scale. Baidu has not published absolute parameter counts.

The official release on May 9, 2026 put it fourth on the LMArena Search leaderboard with a score of 1,223, first among Chinese models, and claims 99.6 on AIME26 with tools.

A preview build had placed 13th on the LMArena text leaderboard on April 30.

On the API it is the shortest-context model of the three: 128K tokens, with 119K of that usable as input and 65,536 tokens of output. Prices step up once the input passes 32K tokens, from CNY 4 and 18 to CNY 6 and 22 per million.

Qianfan lists ERNIE 5.1 under text generation only, not under its deep-thinking models, and Baidu named it as the replacement when it retired ERNIE X1.1. Baidu has released no ERNIE 5.1 weights.

ERNIE-Light-DT-1.0

Light-DT-1.0 is the only ERNIE model Baidu sells in US dollars: $1.13 input and $3.94 output per million tokens on OpenRouter, served by Baidu Qianfan. On Qianfan itself it costs CNY 8 and CNY 28, with cache hits at CNY 0.8.

Baidu's listing says Baidu built it on open-source GLM models and adds Baidu Sparse Attention plus work from Baidu Search's Lightning model.

I found no Baidu release note for it, so treat that description as the only public account of what is inside.

Its selling point is control. It reads 1M tokens, writes up to 128K, and is the one current ERNIE where you can turn thinking off or choose the effort level.

Despite the "Light" name, it costs more per token than 5.5 Flash Preview at today's promo rates.

What the Current ERNIE Models Can Do

All three API models accept tool definitions, support parallel tool calls and return structured JSON through response_format, per Qianfan's default parameter reference.

Qianfan exposes them through an OpenAI-compatible v2 endpoint, and OpenRouter also serves both 5.5 Flash Preview and Light-DT-1.0 through its Anthropic-style messages endpoint.

Multimodal input is the gap. ERNIE 5.0 read text, images, audio and video in one model, and its technical report describes 2.4 trillion parameters with about 3% active per token.

With 5.0 and ERNIE-4.5-Turbo-VL retired from Qianfan, Qianfan lists none of the current ERNIE API models under vision understanding. If you need image input from Baidu today, the open ERNIE 4.5 VL weights are the route.

Consumers meet ERNIE through the chat site at yiyan.baidu.com and the assistant in the Baidu app.

ERNIE 5.0 launched on January 22, 2026, and China Daily reported that it went to individuals through the app and website and to developers through Qianfan, and Baidu's 5.1 release linked to the same chat site.

Baidu does not publish which model version the app runs on a given day. For a hands-on test of the chatbot, see our ERNIE chatbot review.

Context Windows and Output Limits

Two of the three API models read 1M tokens, enough for a mid-sized codebase in one request.

Qianfan lists ERNIE 5.5 Flash Preview with a 1M maximum output as well, against a 64K default for max_tokens; OpenRouter caps the same model's output at 921,600 tokens. Light-DT-1.0 writes up to 131,072 tokens and defaults to 65,536.

ERNIE 5.1 is the constraint: 128K in total, with 119K usable as input.

The jump to 1M is recent. ERNIE X1.1 extended its context to 64K in September 2025, per the Qianfan model update log, and both the ERNIE 4.5 open weights and ERNIE 5.0 stopped at 128K.

Which ERNIE Model Should You Use?

Start on ERNIE 5.5 Flash Preview. It is the cheapest, has the longest context, and Baidu demonstrates it on long coding and agent tasks.

Move to Light-DT-1.0 when you need thinking off for latency, and test ERNIE 5.1 on Chinese-language writing, where Baidu says its internal evaluation puts it close to Gemini 3.1 Pro.

Build on ERNIE With the Right Engineers

Running ERNIE in production means a Qianfan account, retirement notices a month apart and a fallback model ready, which is work for engineers who have shipped on Chinese APIs.

Our LLM developer cost guide puts a mid-level US freelancer at $119 to $185 an hour, against $23 to $39 for a vetted LLM developer from Asia through Second Talent. We send a shortlist within 24 hours.