Approved models
5
Records published from this collection
Provider information
lab
Review approach
Verify each model’s source, terms, capabilities, and current availability
Baidu publishes the ERNIE family for text, vision-language and image generation. These checkpoints need different input pipelines and should be compared within the task they perform.
ERNIE 4.5 21B A3B is the text checkpoint profiled here. ERNIE 4.5 VL 28B A3B adds a vision-language interface. The -PT repository suffix identifies Transformer-style PyTorch weights; it should not be interpreted as a different training stage.
Test English and Chinese documents, ambiguous questions and passages with similar names. For visual inputs, include tiny labels and multi-column layouts. Record language-specific results so one strong subset does not hide an operational weakness.
ERNIE Image and ERNIE Image Turbo are text-to-image releases. Turbo is the distilled route. Compare exact wording, composition and object counts using the publisher’s recommended settings and a shared asset brief.
The checkpoints listed in this refresh use Apache-2.0 licensing. Select the exact repository, tokenizer, preprocessing and runtime before estimating hardware. A conversion or third-party API requires its own evaluation and current terms. Weight parameter counts alone do not capture context memory, batch size and precision.
Review each record individually; collection membership does not imply a recommendation or guarantee of availability.
| ERNIE 4.5 21B A3B Baidu (ERNIE) | 0.0 | — | — | — | — |
| baidu/Unlimited-OCR Baidu (ERNIE) | 5.0 | — | — | — | — |
| ERNIE 4.5 VL 28B A3B Baidu (ERNIE) | 5.0 | — | — | — | — |
| ERNIE Image Baidu (ERNIE) | 5.0 | — | — | — | — |
| ERNIE Image Turbo Baidu (ERNIE) | 5.0 | — | — | — | — |
1–5 of 5 models · select a column to sort
About this order: Recommended shows our editorial shortlist first, followed by recent releases. The spec score is a 0–100 figure SyncDev calculates from each vendor's published specifications — context window, reasoning support, input modalities, tool calling, maximum output and how recently the model shipped. It measures breadth of capability, not benchmark performance, so a higher-scoring model is not automatically the better choice for your task.