Baidu · Text
ERNIE models: Versions, specs, GPU requirements & hosting
Baidu ERNIE 4.5 open MoE checkpoints provide large and compact sibling variants with official PaddlePaddle and FastDeploy support.
Overview and history
ERNIE is an official Baidu model family represented here through independently verified release, checkpoint, license, runtime and hardware evidence. Baidu ERNIE 4.5 open MoE checkpoints provide large and compact sibling variants with official PaddlePaddle and FastDeploy support. Current public releases are grouped by the General, Reasoning, Vision tracks so sibling sizes and specialist branches are not presented as false chronological generations.
FusionPointAI currently tracks 5 certified public ERNIE releases. The catalogue preserves release dates and evaluates Current, Recent, Previous and Legacy status within each model track rather than by database insertion order.
Large and compact MoE sizes
Official FastDeploy path
Long-context text generation
Architecture and model range
The verified public range spans 21B to 424B total parameters and context windows up to 131,072 tokens. Each release states whether it is dense or mixture-of-experts and records active parameters separately when the official source publishes them.
Public releases5
Model tracks3
Current generationERNIE 4.5 VL 28B-A3B Thinking
Last verified2026-09-03T00:00:00.000000Z
Generations, tracks and variants
General
ERNIE 4.5
Current generation 2 sibling variantsReasoning
ERNIE 4.5 Thinking
Current generationVision
Deployment ecosystem
Choose a specific ERNIE release and certified checkpoint before sizing hardware. FusionPointAI rejects unknown runtime configurations, calculates resident memory from total model weights, then distinguishes globally compatible GPUs from currently rentable inventory.
Validate the exact checkpoint license, runtime version, precision kernel, context length, concurrency and multi-GPU topology. Active MoE parameters affect compute, but resident model memory continues to use total stored weights.
Common use cases
Enterprise language services
Chinese and English assistants
PaddlePaddle deployments
License overview
Licenses are certified per release, not inferred from the provider. Review the release page before commercial deployment because terms can differ between generations and variants.
Commercial-use status is a source-backed catalogue classification, not legal advice. Follow the linked official license and attribution requirements.
Release timeline
baidu/ernie/ernie-4-5-300b-a47b @ 454d7e56cd1683c0b6255bb89c949b47f8d8c99c917c849b57a632836d71522a
baidu/ernie/ernie-4-5-21b-a3b @ 55376fc68504404a927791015701e9c8b79b9a55d204aa833581c936fab6aeca
baidu/ernie/ernie-4-5-21b-a3b-thinking @ aa54548a750802c4c7471b17b3a3b850f3f4b0465a2cf4c43f0d1dd9c9dc5297
baidu/ernie/ernie-4-5-vl-28b-a3b-thinking @ a88e2de785ab3000d4b7062ce96d896d2e36ecfa9c8940a9d2da110cc3fba3da
Frequently asked questions
Which ERNIE release is current?
Current releases are ERNIE 4.5 300B-A47B, ERNIE 4.5 21B-A3B, ERNIE 4.5 21B-A3B Thinking, ERNIE 4.5 VL 28B-A3B Thinking, ERNIE 4.5 VL 424B-A47B, evaluated within their respective tracks.
Can ERNIE be self-hosted?
Only releases with a verified official checkpoint and supported runtime configuration are presented as deployable.
Which GPU should run ERNIE?
GPU selection depends on the exact variant, checkpoint precision, context length and concurrency; use the release hardware recommendations rather than the family name alone.
Official references
Last verified: 2026-09-03