B

Baidu: ERNIE 4.5 21B A3B Thinking

Baidu
Vision
131K
Context
$0.0600
Input /1M
$0.2200
Output /1M
N/A
Max Output
Run API Call Open in Chat

Baidu: ERNIE 4.5 21B A3B Thinking

Model ID: baidu/ernie-4.5-21b-a3b-thinking Provider: Baidu (via NovitaAI) Category: Reasoning Model, Multilingual Release Date: October 9, 2025 Parameters: 21 billion

Overview

ERNIE-4.5-21B-A3B-Thinking is Baidu's upgraded lightweight MoE (Mixture of Experts) model refined to boost reasoning depth and quality. It delivers top-tier performance in logical puzzles, math, science, coding, text generation, and expert-level academic benchmarks. The model excels across multilingual reasoning tasks.

Technical Specifications

Property Value
Parameters 21 billion
Architecture Mixture of Experts (MoE)
Context Length 131,072 tokens
Input Modalities Text
Output Modalities Text
Max Completion Tokens 65,536
Reasoning Format <think></think> tags

Pricing

Via NovitaAI Provider:

Type Price
Input $0.06 per 1M tokens
Output $0.22 per 1M tokens

Quantization: Standard

Capabilities

Reasoning & Analysis

  • Deep reasoning with thinking tokens
  • Logical puzzle solving
  • Mathematical problem solving
  • Scientific reasoning
  • Code understanding and generation
  • Text generation and summarization
  • Expert-level academic benchmarks

Input Modalities

  • Text only

Output Modalities

  • Text only

Key Features

  • Reasoning support with configurable depth
  • Reasoning output wrapped in <think></think> tags
  • Mixture of Experts architecture for efficiency
  • Lightweight 21B parameters
  • Extended context window (131,072 tokens)
  • Structured outputs support

Supported Parameters

  • reasoning - Enable/configure reasoning mode
  • include_reasoning - Include reasoning in output
  • max_tokens - Maximum output tokens
  • temperature - Sampling temperature control
  • top_p - Nucleus sampling parameter
  • stop - Stop sequences for output termination
  • frequency_penalty - Reduce repetitive tokens
  • presence_penalty - Encourage diverse content
  • seed - Random seed for reproducibility
  • top_k - Top-K sampling parameter
  • repetition_penalty - Control repetition

Use Cases

  • Mathematical Problem Solving: Complex math and logic puzzles
  • Scientific Analysis: Scientific reasoning and research support
  • Code Generation: Programming and code understanding
  • Academic Research: Expert-level analysis and writing
  • Multilingual Reasoning: Reasoning across multiple languages
  • Technical Writing: Documentation and technical content
  • Logic Puzzles: Complex logical reasoning challenges
  • Coding Interviews: Preparation and problem solving

Limitations

  • Text Only: No image or video understanding
  • Reasoning Output: Must enable reasoning explicitly
  • Moderate Context: 131K tokens (smaller than some alternatives)

Best Practices

  1. Enable Reasoning: Always enable reasoning for complex tasks
  2. Leverage MoE: Architecture is optimized for efficiency
  3. Temperature Setting: Lower temperature (0.3-0.7) for logical tasks
  4. Context Usage: Use full 131K token context for long documents
  5. Multilingual: Good for applications requiring cross-language reasoning
  • DeepSeek: DeepSeek V3.2 - Alternative reasoning model
  • AllenAI: Olmo 3.1 32B Think - Free reasoning alternative
  • Google: Gemini 3 Flash Preview - Multimodal reasoning
  • Anthropic: Claude 3.5 Sonnet - Alternative reasoning model

Performance Metrics

Benchmark Performance

  • Top-tier performance on:
    • Logical Puzzles: Advanced logic and reasoning tasks
    • Mathematics: Complex mathematical problem solving
    • Science: Scientific reasoning and analysis
    • Coding: Code generation and understanding
    • Academic Tasks: Expert-level academic benchmarks

Usage Statistics (Recent)

  • Consistent daily usage patterns
  • Variable request volumes indicating adoption for reasoning-intensive tasks
  • Strong performance across multilingual inputs

Provider Information

Primary Provider: NovitaAI

Data Policy

  • Training Use: Not used for model training
  • Prompt Retention: Prompts not retained
  • Publishing: Cannot publish outputs without permission

Multilingual Support

ERNIE 4.5 excels at reasoning across multiple languages:

  • Chinese: Native performance
  • English: Proficient
  • Other Languages: Strong multilingual reasoning

Advantages

Efficiency

  • 21B Parameters: Lightweight compared to larger reasoning models
  • MoE Architecture: Efficient expert routing
  • Cost-Effective: Competitive pricing for reasoning capability

Performance

  • Expert-Level Benchmarks: Top performance on academic tasks
  • Reasoning Depth: Strong logical and mathematical reasoning
  • Multilingual: Excellent cross-language reasoning

Value

  • Balanced Pricing: Mid-range cost for strong reasoning
  • Versatile: Works well across multiple domains

Comparison with Alternatives

Model Parameters Reasoning Multimodal Cost (Input)
ERNIE 4.5 21B 21B Yes No $0.056/M
DeepSeek V3.2 Large Yes No $0.224/M
Gemini 3 Flash Unknown Yes Yes $0.50/M
Claude 3.5 Sonnet 200B+ Yes Yes $3/M

Output Format

Reasoning Output

When reasoning is enabled, the model returns output in this format:

<think>
[Internal reasoning chain and logic]
</think>

[Final answer based on reasoning]

Additional Notes

  • Baidu Research: Backed by Baidu's strong AI research capabilities
  • Lightweight Design: 21B parameters make it suitable for varied deployments
  • Competitive Pricing: Excellent value for reasoning capabilities
  • MoE Benefits: Mixture of Experts provides capability without size penalty
  • Emerging Strong: Growing adoption for reasoning-focused applications
  • Quality Assurance: Expert-level benchmark performance validates quality

Training & Data

  • Trained on 9 trillion tokens (inherited from ERNIE base)
  • Focus on reasoning and instruction following
  • Academic and technical domain emphasis
  • Multilingual training dataset

Recommendation Scenarios

Ideal for:

  • Cost-conscious reasoning applications
  • Multilingual reasoning tasks
  • Academic and research applications
  • Code generation and analysis
  • Technical problem solving

Consider alternatives for:

  • Vision/multimodal requirements
  • Maximum reasoning depth (use larger models)
  • Real-time high-volume applications (check rate limits)