ERNIE 3.5

What it is

ERNIE 3.5 is a large language model from the Chinese search company Baidu, announced in late June 2023. It was not launched as a stand-alone product: Baidu presented it as the upgraded model running behind ERNIE Bot, its chatbot. Up to that point ERNIE Bot had been available only through invitation-only testing, running since March on the earlier ERNIE 3.0 model.^[1][2]

The name follows Baidu's ERNIE family of models, which the company develops for Chinese-language tasks and integrates into its search and cloud products.

How it arrived

The upgrade surfaced in the chatbot first. On 21 June 2023 Baidu released ERNIE Bot version 2.1.0, which added the ChatFile plugin and improved the bot's mathematics and creative writing, with ERNIE 3.5 credited for those gains.^[1] ChatFile lets a user hand a long document to the model and get a summary back, rather than working through it passage by passage.

The release also introduced a plugin architecture. The first two plugins were a Baidu Search plugin, which pulls in current information rather than relying only on what the model learned during training, and the ChatFile plugin. Baidu said it planned to open the plugin system to third parties.^[1]

What Baidu claimed

Baidu said ERNIE 3.5 surpassed ChatGPT — by which it meant GPT-3.5 — on comprehensive ability scores, and beat GPT-4 on several Chinese-language capabilities. The company cited tests carried out by the state newspaper China Science Daily using two benchmarks, AGIEval and C-Eval.^[1]

Those figures were Baidu's own account of the comparison, and they describe aggregate scores and Chinese-language tasks rather than a single overall ranking. Benchmarks of this kind measure a model on a fixed set of questions; they say something about how a model handles test-style problems, and less about how it behaves in open-ended, everyday use.^[1]

Efficiency

Beyond quality, Baidu framed the release as an engineering step. The company reported roughly twice the training throughput and seventeen times the inference throughput of the previous version.^[1] In plain terms, training throughput is how much data the model can process per unit of time while it is being built; inference throughput is how many responses it can serve once deployed to users.

Baidu said the improved training and inference efficiency would allow faster and cheaper future iterations of the model.^[2] That matters commercially: a chatbot that answers many users costs money on every reply, so serving efficiency feeds directly into what a provider can afford to offer and at what price.

Where it sat in the line-up

ERNIE 3.5 was a waypoint rather than an endpoint. It was succeeded by ERNIE 4.0, which Baidu unveiled on 17 October 2023.^[5] When the paid ERNIE 4.0 plan started in November 2023, the basic ERNIE 3.5 version of ERNIE Bot remained free — a two-tier arrangement in which the older model carried the free user base and the newer one carried the subscription.^[3]

The model stayed available to developers after the newer release. As of April 2024 it was offered through Baidu's ModelBuilder platform alongside ERNIE 4.0, and Baidu described ModelBuilder as curating up to 77 models across different sizes, drawn from ERNIE 3.5 and 4.0 among others.^[4]

Why it mattered

ERNIE 3.5 marks the point at which Baidu moved its chatbot from closed testing to a broadly available product with an explicit feature set — plugins, document summarisation, live search — and began arguing in public that its models matched or beat OpenAI's on measured tasks.^[1][2] It also shows the rhythm of the field at the time: within roughly four months the company had announced a successor, and the previous generation was pushed down to the free tier.^[3][5]