ERNIE 4.5 300B A47B is Baidu's flagship open-weight entry in the ERNIE 4.5 series, built on a heterogeneous Mixture-of-Experts architecture that totals around 300 billion parameters while activating roughly 47 billion per token. According to its OpenRouter listing, the model is fine-tuned for language-only tasks and uses advanced routing along with FP8 and 2-bit quantization strategies to keep inference efficient at that scale. The full pretrained checkpoint is published on Hugging Face under baidu/ERNIE-4.5-300B-A47B-PT, giving researchers and developers direct access to the weights rather than relying solely on hosted endpoints.
In practice, the model is positioned for high-throughput general-purpose applications that demand strong reasoning alongside long-context handling. OpenRouter describes support for English and Chinese text generation, tool parameters, and extended contexts up to 131K tokens, with a reported release on June 30, 2025 and a knowledge cutoff in March 2025. AIHubMix's comparison page shows it priced at $0.32 per million input tokens and $1.28 per million output tokens through that aggregator, though rates vary across routing providers. The combination of sparse activation, open weights, and a very large context window makes it a flexible choice for multilingual assistants, long-document analysis, and tool-augmented workflows where both depth of reasoning and throughput matter.
