Tencent: Hunyuan A13B Instruct

Hunyuan-A13B uses a sparse Mixture-of-Experts (MoE) architecture with 13B active parameters selected from an 80B parameter pool, enabling efficient instruction-following through dynamic expert routing. The model supports explicit chain-of-thought reasoning patterns, allowing it to decompose complex tasks into intermediate reasoning steps before generating final responses. This architecture reduces computational overhead during inference while maintaining reasoning capability through selective expert activation based on input tokens.

Hunyuan-A13B is instruction-tuned to follow multi-turn conversational patterns, maintaining coherence across sequential user requests within a single session. The model processes each turn as context-aware input, allowing it to reference previous exchanges and adapt responses based on conversation history. This capability enables natural dialogue flows where the model understands implicit references, maintains consistent persona, and refines answers based on user feedback across turns.

Hunyuan-A13B can generate code snippets and provide technical explanations by leveraging its instruction-tuning and chain-of-thought capability. When prompted with code-related tasks, the model can produce syntactically valid code in multiple languages, explain implementation logic, and reason through algorithmic problems. The MoE architecture may route to specialized experts for code understanding, though this is implementation-dependent and not explicitly documented.

Hunyuan-A13B is designed to achieve competitive performance on standard instruction-following benchmarks (MMLU, HellaSwag, TruthfulQA, etc.) through instruction-tuning and MoE specialization. The model's architecture allows different experts to specialize in different task domains, enabling strong cross-domain performance without proportional parameter scaling. This capability reflects the model's training on diverse instruction datasets and evaluation against established baselines.

Hunyuan-A13B is accessible via OpenRouter's API, providing a managed inference endpoint without requiring local deployment or infrastructure management. The integration handles model loading, batching, and scaling transparently, exposing a standard REST API interface for text generation. Developers interact with the model through HTTP requests, specifying parameters like temperature, max tokens, and top-p sampling, with responses streamed or returned in full depending on configuration.

Hunyuan-A13B supports streaming generation through OpenRouter's API, allowing responses to be consumed token-by-token as they are generated rather than waiting for full completion. This capability enables real-time user feedback, progressive rendering in UIs, and early stopping based on application logic. The model exposes sampling parameters (temperature, top-p, top-k) for fine-grained control over generation behavior, allowing tuning of output diversity and determinism.