OpenAI: gpt-oss-120b

Implements a 117B-parameter Mixture-of-Experts architecture that activates only 5.1B parameters per forward pass, routing input tokens to specialized expert subnetworks based on learned gating functions. This sparse activation pattern reduces computational cost while maintaining model capacity for complex reasoning tasks, using a load-balancing mechanism to distribute tokens across experts and prevent collapse to a single dominant expert.

Supports structured reasoning chains where the model can decompose complex tasks into intermediate steps, make decisions about which tools or functions to invoke, and iteratively refine outputs based on tool results. The model is trained to generate reasoning tokens that explicitly show its decision-making process, enabling transparent multi-turn agent loops where each step's output feeds into the next step's input, with native support for function calling schemas and structured output formatting.

Processes up to 128,000 tokens in a single context window, enabling the model to maintain coherent understanding across entire documents, codebases, or multi-turn conversations without losing semantic relationships between distant parts of the input. Uses efficient attention mechanisms (likely sparse or linear attention variants optimized for MoE) to handle long sequences while maintaining the reasoning capability needed for complex analysis across the full context.

Generates syntactically correct and semantically sound code across 40+ programming languages (Python, JavaScript, Java, C++, Go, Rust, etc.), with understanding of language-specific idioms, frameworks, and best practices. The model is trained on diverse code repositories and can generate complete functions, classes, or multi-file solutions, with support for generating code that integrates with popular libraries and frameworks. Includes capability to understand existing code context and generate compatible additions or refactorings.

Reliably follows complex, multi-part instructions and generates output in specified structured formats (JSON, XML, YAML, CSV, Markdown tables) with high consistency. The model is trained to parse instruction hierarchies, handle conditional logic (if-then patterns), and generate output that strictly adheres to specified schemas or templates. Supports both explicit format requests (e.g., 'output as JSON') and implicit format inference from examples provided in the prompt.

Provides inference through OpenAI's REST API with support for both streaming (real-time token-by-token output) and batch processing (asynchronous processing of multiple requests). Streaming mode returns tokens as they are generated, enabling real-time user feedback and progressive rendering in applications. Batch mode accepts multiple requests in a single API call, optimizing throughput for non-latency-sensitive workloads and reducing per-request overhead through request consolidation.

Understands and generates text in 50+ languages with reasonable fluency, including major languages (Spanish, French, German, Mandarin, Japanese, Arabic) and many lower-resource languages. The model maintains semantic understanding across language boundaries and can perform tasks like translation, cross-lingual information retrieval, and multilingual summarization. Uses language-agnostic tokenization and embedding spaces to handle diverse character sets and linguistic structures.

Maintains coherent conversation state across multiple turns, where each response is informed by the full conversation history and previous context. The model tracks entities, relationships, and discussion topics across turns, enabling natural follow-up questions and references to earlier statements without explicit re-specification. Uses attention mechanisms to weight recent context more heavily while still maintaining awareness of earlier conversation points, with support for explicit context management through system prompts and conversation summaries.

Model has a training data cutoff date (typically April 2024 or later based on OpenAI's release patterns) and is aware of its knowledge limitations. The model can acknowledge when information is outside its training data and can be prompted to reason about recent events using provided context. Does not have real-time internet access but can be augmented with retrieval-augmented generation (RAG) systems to access current information.