Baidu: ERNIE 4.5 21B A3B

Generates text using a 21B parameter Mixture-of-Experts architecture that activates only 3B parameters per token through learned routing mechanisms. This sparse activation pattern reduces computational overhead while maintaining model capacity, using heterogeneous expert specialization where different experts handle distinct semantic or linguistic domains. The routing mechanism learns to select which expert subset processes each token based on input context.

Processes both text and image inputs through a unified architecture where modality-isolated routing directs image and text tokens to specialized expert subsets. The model encodes images into token sequences (likely through a vision encoder) and routes them through experts trained specifically for visual understanding, while text tokens follow separate routing paths. This heterogeneous design allows the model to reason across modalities without forcing all experts to handle both equally.

Maintains conversation state across multiple turns by accepting full conversation history in API requests and using attention mechanisms to track context dependencies. The model processes the entire conversation history to generate contextually appropriate responses, with routing decisions informed by prior turns. This approach allows the model to reference earlier statements, maintain consistent character or tone, and resolve pronouns and references across turns.

Generates text incrementally through token-by-token streaming, allowing clients to receive and display partial responses before generation completes. The API returns tokens as they are generated rather than waiting for full completion, enabling real-time user feedback and lower perceived latency. This is implemented through HTTP streaming (likely Server-Sent Events or chunked transfer encoding) where each token is sent as it exits the sparse MoE routing and generation pipeline.

Exposes the ERNIE 4.5 21B model through OpenRouter's unified API interface, allowing developers to call the model using standard HTTP requests without direct Baidu API integration. OpenRouter handles authentication, rate limiting, and request routing, providing a consistent interface across multiple model providers. Requests are formatted as JSON with standard chat completion schemas, and responses follow OpenAI-compatible formats for easy integration with existing LLM tooling.

Reduces inference costs by activating only 3B of 21B parameters per token, lowering computational requirements and memory bandwidth compared to dense models. The sparse activation is achieved through learned routing that selects which expert subset processes each token based on input content. This architectural choice reduces floating-point operations (FLOPs) and memory access patterns, directly translating to lower API costs and faster inference latency.