Baidu: ERNIE 4.5 300B A47B

ERNIE-4.5-300B-A47B implements a Mixture-of-Experts (MoE) architecture where only 47B out of 300B total parameters are activated per token, reducing computational overhead while maintaining model capacity. The model uses a gating network to route tokens to specialized expert modules, enabling efficient inference through sparse activation patterns rather than dense forward passes through all parameters.

ERNIE-4.5-300B-A47B processes conversation history through explicit system/user/assistant message roles, maintaining coherent context across multiple exchanges without requiring manual context window management. The model implements sliding-window attention or similar context compression to handle extended dialogues while respecting token limits, enabling stateless API calls where conversation state is passed in each request.

ERNIE-4.5-300B-A47B is trained on instruction-following datasets enabling it to interpret natural language task descriptions and adapt behavior accordingly. The model uses in-context learning to follow complex multi-step instructions, system prompts for behavioral constraints, and few-shot examples to guide output format — all without fine-tuning, leveraging the model's learned ability to parse and execute arbitrary instructions.

ERNIE-4.5-300B-A47B supports text generation across multiple languages (Chinese, English, and others) through language-agnostic MoE routing where the gating network treats tokens uniformly regardless of language, allowing the model to leverage shared expert knowledge across linguistic boundaries. The model was trained on multilingual corpora, enabling code-switching and cross-lingual reasoning without language-specific model variants.

ERNIE-4.5-300B-A47B is accessed exclusively via OpenRouter or Baidu's API, supporting both streaming (token-by-token output for real-time UI) and batch (full completion returned at once) inference modes. The API abstracts away model deployment complexity, handling load balancing, rate limiting, and multi-user concurrency server-side, while clients manage request formatting and response parsing.

ERNIE-4.5-300B-A47B exposes temperature, top-p (nucleus sampling), and top-k parameters allowing fine-grained control over output randomness and diversity. Lower temperatures (0.0-0.5) produce deterministic, focused outputs suitable for factual tasks; higher temperatures (0.7-1.0+) increase creativity and diversity for open-ended generation. The model implements standard softmax temperature scaling and nucleus sampling, enabling developers to tune the probability distribution over tokens without retraining.

ERNIE-4.5-300B-A47B allows clients to specify max_tokens parameter, controlling the maximum length of generated completions. This enables developers to enforce output length constraints without post-processing, useful for fitting responses into UI constraints or limiting API costs. The model respects the max_tokens limit during generation, stopping early if the limit is reached before natural completion.

ERNIE-4.5-300B-A47B supports stop_sequences parameter allowing developers to specify custom tokens or strings that trigger generation termination. When the model generates a stop sequence, output is immediately halted and returned, enabling natural conversation boundaries (e.g., stopping at newlines for single-line outputs) or domain-specific delimiters without post-processing.