Upstage: Solar Pro 3

Solar Pro 3 implements a Mixture-of-Experts (MoE) architecture with 102B total parameters but only activates 12B parameters per forward pass through learned gating mechanisms that route tokens to specialized expert subnetworks. This selective activation pattern reduces computational cost while maintaining model capacity, using sparse expert selection rather than dense transformer layers for each token position.

Solar Pro 3 maintains conversation state across multiple turns by accepting full conversation history in each API request, with support for extended context windows that allow retention of longer dialogue histories and document context. The model processes the entire conversation context through its MoE routing mechanism, enabling coherent multi-turn interactions without explicit memory management.

Solar Pro 3 generates syntactically correct code across multiple programming languages (Python, JavaScript, Java, C++, SQL, etc.) by leveraging its 102B parameter capacity trained on diverse code corpora. The MoE architecture routes code-generation tokens to specialized experts trained on language-specific patterns, enabling context-aware completions that respect language idioms and frameworks.

Solar Pro 3 accepts system prompts that define behavioral constraints and task-specific instructions, then follows those instructions consistently across multiple turns. The model decomposes complex tasks into subtasks by analyzing the system prompt and user request, routing different reasoning steps through appropriate expert pathways in its MoE architecture.

Solar Pro 3 performs semantic analysis and reasoning by processing input text through its 102B parameter capacity, with MoE routing directing reasoning-heavy tokens to expert subnetworks trained on logical inference and knowledge synthesis. The model can answer questions requiring multi-step reasoning, identify semantic relationships, and synthesize information across multiple concepts.

Solar Pro 3 supports streaming inference through OpenRouter's API, returning tokens incrementally as they are generated rather than waiting for the complete response. This enables real-time display of model output in user interfaces, reducing perceived latency and allowing users to see reasoning progress as it unfolds.

Solar Pro 3 is accessed exclusively through OpenRouter's REST API, accepting configuration parameters like temperature, top-p, top-k, and max-tokens to control output randomness and length. The API abstracts away model deployment complexity, handling load balancing and infrastructure while exposing a simple HTTP interface for inference requests.

Solar Pro 3 generates original content across multiple genres and styles (marketing copy, creative fiction, technical documentation, etc.) by conditioning on style descriptors and examples in prompts. The model's 102B parameters provide sufficient capacity for diverse writing styles, with MoE routing allowing different experts to specialize in different genres.