技术规格

GPT-OSS 20B

GPT-OSS 20B is a text-based language model developed by OpenAI, specifically engineered to deliver high-performance reasoning on consumer-grade hardware. As part of the GPT-OSS family, this model balances computational efficiency with complex task execution, utilizing a sparse architecture to maintain a low memory footprint. It is designed to function as a flexible component in local and enterprise environments, where data privacy and low-latency response times are critical requirements.

The model utilizes a Mixture-of-Experts (MoE) transformer architecture consisting of 24 layers. While the total parameter count is 21 billion, the system only activates 3.6 billion parameters per token during the forward pass. This sparsity is achieved through a routing mechanism that selects four active experts from a pool of 32 for each token. The architecture incorporates several modern optimizations, including SwiGLU activation functions, Root Mean Square (RMS) normalization, and Grouped-Query Attention (GQA) with eight key-value heads to optimize memory throughput. It also supports a native context window of 128,000 tokens using Rotary Positional Embeddings (RoPE).

Functionally, GPT-OSS 20B is optimized for agentic workflows and complex reasoning tasks. It supports features such as native tool use, function calling, and a configurable reasoning effort system that allows developers to adjust the model's processing depth based on the specific latency needs of the application. The model is trained using a specialized response format to facilitate consistent structured outputs and long-form chain-of-thought reasoning, making it suitable for scientific analysis, code generation, and specialized technical assistance on local devices.