Baidu: ERNIE 4.5 21B A3B

Text input Text output
Author's Description

A sophisticated text-based Mixture-of-Experts (MoE) model featuring 21B total parameters with 3B activated per token, delivering exceptional multimodal understanding and generation through heterogeneous MoE structures and modality-isolated routing. Supporting an extensive 131K token context length, the model achieves efficient inference via multi-expert parallel collaboration and quantization, while advanced post-training techniques including SFT, DPO, and UPO ensure optimized performance across diverse applications with specialized routing and balancing losses for superior task handling.

Key Specifications
Cost
$$
Context
120K
Parameters
21B
Released
Aug 12, 2025
Speed
Ability
Reliability
Supported Parameters

This model supports the following parameters:

Logit Bias Stop Seed Min P Top P Max Tokens Frequency Penalty Temperature Presence Penalty
Performance Summary

Baidu's ERNIE 4.5 21B A3B demonstrates competitive response times, ranking in the 56th percentile for speed across various benchmarks. Its pricing is a significant advantage, consistently placing it among the most competitive models with an 81st percentile ranking. The model exhibits exceptional reliability, boasting a 98% success rate, indicating minimal technical failures. In terms of performance across categories, ERNIE 4.5 21B A3B shows strong capabilities in Hallucinations (92.0% accuracy) and General Knowledge (94.3% accuracy), suggesting a good understanding of factual information and an ability to acknowledge uncertainty. It also performs well in Ethics (94.5% accuracy). However, its accuracy in Mathematics (82.0%), Email Classification (93.5%), Instruction Following (51.2%), and Reasoning (50.0%) falls into the lower percentiles, indicating areas for potential improvement. Coding performance is moderate at 77.2%. Key strengths include its high reliability and cost-effectiveness, making it an attractive option for applications where these factors are paramount. Its primary weaknesses lie in complex reasoning and instruction following, where accuracy is comparatively lower.

Model Pricing

Current Pricing

Feature Price (per 1M tokens)
Prompt $0.07
Completion $0.28

Price History

Available Endpoints
Provider Endpoint Name Context Length Pricing (Input) Pricing (Output)
Novita
Novita | baidu/ernie-4.5-21b-a3b 120K $0.07 / 1M tokens $0.28 / 1M tokens
Benchmark Results
Benchmark Category Reasoning Strategy Free Executions Accuracy Cost Duration
Other Models by baidu