OpenAI: gpt-oss-20b

Executes forward passes using a Mixture-of-Experts (MoE) architecture where only 3.6B of 21B parameters are active per token, routing each token to specialized expert sub-networks via learned gating functions. This sparse activation pattern reduces computational cost and memory bandwidth compared to dense models while maintaining parameter capacity for diverse reasoning tasks.

Maintains coherent multi-turn dialogue by processing conversation history within a fixed context window, using attention mechanisms to weight recent and relevant prior messages while discarding or summarizing older context when token limits are approached. The model learns to extract key information from conversation history to maintain semantic continuity across turns.

Generates syntactically valid code across multiple programming languages by learning patterns from training data that includes code repositories, technical documentation, and problem-solution pairs. The model applies language-specific reasoning to produce working implementations, debug explanations, and architectural suggestions for technical problems.

Answers factual and conceptual questions by retrieving and synthesizing relevant knowledge from training data, applying reasoning to connect concepts across domains. The model generates coherent explanations that cite reasoning steps and provide context-appropriate detail levels based on question complexity.

Interprets complex, multi-step instructions and decomposes them into executable sub-tasks, then generates outputs following specified constraints (format, length, tone, structure). The model learns to parse instruction syntax, identify priorities, and handle edge cases like conflicting constraints or ambiguous requirements.

Generates original creative content (stories, poetry, marketing copy, dialogue) by learning stylistic patterns, narrative structures, and genre conventions from training data. The model applies learned constraints (rhyme schemes, character consistency, tone) to produce coherent creative outputs that match specified requirements.

Condenses long-form text into concise summaries by identifying key information, removing redundancy, and preserving essential meaning. The model learns to extract structured information (entities, relationships, facts) from unstructured text and present it in specified formats (bullet points, JSON, tables).

Translates text between languages and generates content in non-English languages by learning multilingual patterns from training data. The model preserves meaning, tone, and context-appropriate phrasing across language pairs, and can switch between languages within a single response.

Solves mathematical problems and performs logical reasoning by learning to apply mathematical rules, algebraic manipulation, and logical inference patterns from training data. The model generates step-by-step solutions, explains reasoning, and handles problems ranging from arithmetic to calculus and symbolic logic.

Exposes model inference through OpenRouter's API, providing OpenAI-compatible endpoints that accept standard chat completion requests and return structured responses. Integration handles authentication, rate limiting, request routing, and response formatting without requiring direct model deployment.