GPT-OSS employs a Mixture-of-Experts (MoE) architecture with two model variants: gpt-oss-120b contains 117B total parameters with 5.1B active parameters, while gpt-oss-20b has 21B total parameters with 3.6B active parameters. The MoE design is crucial for efficiency, as it allows the models to maintain high performance while only activating a fraction of their parameters during inference, resulting in faster processing and lower memory requirements compared to dense models of similar capability.
The models feature native MXFP4 quantization specifically applied to the MoE layer weights. This 4-bit quantization scheme is a significant technical innovation that enables gpt-oss-120b to run on a single H100 GPU and gpt-oss-20b to run within 16GB of memory. The quantization only affects the MoE weights while keeping other tensors in BF16 format, striking a balance between efficiency and performance. This approach makes the models particularly accessible for deployment on consumer hardware and edge devices.
The reasoning capabilities are among the most advanced features of GPT-OSS. The models support configurable reasoning effort levels (low, medium, high) that can be adjusted based on specific use cases and latency needs. They provide full chain-of-thought access, giving complete visibility into the model’s reasoning process for easier debugging and increased trust in outputs, though this isn’t intended to be shown to end users. Both models were trained on OpenAI’s harmony response format and include native agentic capabilities for function calling, web browsing, Python code execution, and Structured Outputs. The reasoning system allows developers to scale computational effort based on task complexity, making the models versatile for everything from simple queries to complex problem-solving scenarios.
For more detailed information, see: GPT-oss vs o4-mini: Edge-Ready, On-Par Performance — Dependable, Not Mind-Blowing