baidu/ernie-4.5-300b-a47b-paddle

ERNIE 4.5 300B A47B is a super-large-scale Mixture-of-Experts (MoE) language model launched by Baidu, primarily positioned as the “most powerful Chinese MoE architecture and an enterprise-grade intelligent foundation for all scenarios.”

• Ultra-large MoE scale: With a total of 300 billion parameters and approximately 47 billion activated parameters, it is natively trained using PaddlePaddle and specially optimized for Chinese-language scenarios.
• Leading Chinese understanding capabilities: It significantly outperforms comparable open-source and closed-source models in authoritative Chinese benchmarks such as CMMLU and C-Eval.
• Enhanced multilingual capabilities: In addition to Chinese, it supports dozens of languages including English, Japanese, and Korean, meeting the needs of global business operations.
• Long-context support: It natively supports context lengths of up to 128,000 tokens, making it ideal for long-document analysis, multi-turn conversations, and complex reasoning tasks.
• Enterprise-grade toolchain integration: Paired with the ERNIE Bot platform, it supports function calling, knowledge-base enhancement, agent frameworks, and private deployment options.

───────────────────────────────────────────────────────────────────

Core Capabilities

🇨🇳 Deep semantic modeling for Chinese: Accurately captures idioms, classical texts, industry-specific terminology, and local cultural contexts, generating content that better aligns with Chinese user preferences.

🧠 Advanced logic and coding abilities: Demonstrates stable performance in tasks such as mathematical reasoning, algorithm design, and SQL generation, supporting complex engineering scenarios.

🧩 Native agent architecture: Can autonomously call APIs, parse web pages, and interact with databases, enabling the construction of automated workflows for finance, government services, and customer support.

🛡️ Secure, compliant, and controllable: Equipped with built-in content filtering, sensitive-word detection, and data anonymization mechanisms, it meets both domestic regulatory requirements and enterprise security standards.