Tasks and inputs
ERNIE 4.5 21B A3B is listed for Text generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Oct 13, 2026
ERNIE 4.5 21B A3B is a downloadable checkpoint in Baidu's ERNIE 4.5 mixture-of-experts family.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
ERNIE 4.5 21B A3B is a downloadable checkpoint in Baidu's ERNIE 4.5 mixture-of-experts family.
This listing describes ERNIE-4.5-21B-A3B-PT. The text and vision-language releases have different input interfaces. A shared family name does not make their tokenizer, preprocessing or deployment configuration interchangeable.
Use English and Chinese document questions, extraction and instruction-following tasks. Keep language coverage visible in the test results instead of averaging away a weak language.
Use prompts with an answer you can check, along with cases that lack enough evidence. Score the final answer, formatting and the amount of manual correction needed. A confident response should not receive credit merely for sounding complete.
Keep the original repository, revision, tokenizer and runtime versions with the evaluation. When selecting a quantized conversion, inspect the converter's files and test it separately from the publisher checkpoint. Total and active parameter counts describe different aspects of an MoE deployment; neither establishes the complete hardware requirement.
This page records the weight licence from the publisher metadata and leaves managed API pricing unset. A third-party inference host has its own identifier, context allowance and service contract. Confirm those separately when choosing a route.
For a text assistant, compare another instruction-tuned text model on the same documents. For a visual assistant, keep image decoding and preprocessing fixed. These controls make a comparison useful beyond a table of model names.
Publisher source 1 Deployment limits and commercial terms belong to the selected checkpoint or API route.
Compare the model on the work you need to complete.
ERNIE 4.5 21B A3B is listed for Text generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Anthropic
Claude’s small model for high-volume extraction, classification and routing, with adaptive thinking and a 1M-token context window.
Mistral AI
Mistral Large 4 is an open-weight multimodal model released in public preview on the publisher’s announced release, for general reasoning, coding and agent applications.
OpenAI
GPT-6.1 Sol is OpenAI's updated Sol model for coding and professional tasks. It accepts text and images and returns text.
inclusionAI
Ling 3.1 Flash is inclusionAI’s hybrid reasoning model for coding and tool agents, available through a hosted AI Gateway route.
Anthropic
Claude Sonnet 5.5 is an Anthropic language model that accepts text and images and returns text.
MiniMax
MiniMax M3.1 Flash Preview is a multimodal reasoning model for code and agent workflows, accessed through eligible MiniMax plans and tools.