ERNIE 4.5 21B A3B Base PT

Name: ERNIE 4.5 21B A3B Base PT
Author: Baidu

by BaiduERNIE

Open WeightModel weights are publicly available — can be downloaded and self-hosted

Released June 2025131K context≈ 98,304 words21B params

A Chinese-English bilingual specialist with a lean active parameter footprint — only 3 billion parameters fire per forward pass despite a 21 billion total parameter pool. This mixture-of-experts architecture means it handles everyday language tasks efficiently without burning through compute. Its base pretrained nature means it arrives without instruction tuning, so it responds to raw text continuation rather than conversational prompts.

Capabilities

Capability scores are AI-generated based on model documentation, benchmarks, and technical specifications. Learn more

Multimodal

Moderate

Coding

ERNIE 4.5 21B A3B Base PT

by BaiduERNIE

Open WeightModel weights are publicly available — can be downloaded and self-hosted

Released June 2025131K context≈ 98,304 words21B params

A Chinese-English bilingual specialist with a lean active parameter footprint — only 3 billion parameters fire per forward pass despite a 21 billion total parameter pool. This mixture-of-experts architecture means it handles everyday language tasks efficiently without burning through compute. Its base pretrained nature means it arrives without instruction tuning, so it responds to raw text continuation rather than conversational prompts.

Capabilities

Capability scores are AI-generated based on model documentation, benchmarks, and technical specifications. Learn more

Multimodal

Moderate

Coding

Glossary

ArchitectureThe underlying structural design of a neural network that defines how data flows through layers and components.Base PretrainedA model trained only on raw text prediction without additional instruction-following training, so it completes text continuations rather than answering questions or following commands.BilingualA model trained to understand and generate text in two languages, in this case Japanese and English.Forward PassA single computation cycle where input data flows through the model's layers to produce an output prediction.Parameter FootprintThe total number of learnable weights in a model, which directly affects its memory requirements and computational cost — smaller footprints run faster on consumer devices.Parameter PoolThe total set of learnable weights in a model; in sparse models, only a subset of this pool is activated for any given input.ParametersThe learned numerical values in a model — more parameters generally means more capacity but higher compute cost.PretrainedA model that has already been trained on large amounts of text data before being released or fine-tuned for specific tasks.Text ContinuationThe task of generating the next words or sentences based on a given prompt or partial text.

Capabilities

Capabilities

Use Case Fit

Glossary