Introducing gpt-oss

OpenAI has released gpt-oss-120b and gpt-oss-20b, open-weight language models that boast powerful reasoning performance and low cost. Provided under the Apache 2.0 license, these models were trained by combining the techniques of state-of-the-art frontier models like o3 with reinforcement learning (RL), and were optimized to run efficiently even on consumer hardware.

gpt-oss-120b matches OpenAI o4-mini performance on core reasoning benchmarks, and runs efficiently on a single 80GB GPU. gpt-oss-20b delivers performance similar to o3-mini while being capable of running on edge devices with just 16GB of memory, making it ideal for local inference and rapid iteration. Both models have demonstrated excellent performance in tool use, Chain-of-Thought (CoT) reasoning, and HealthBench.

These models are compatible with the Responses API, and exhibit excellent instruction-following ability within agentic workflows such as web search or Python code execution. They also support Structured Outputs, and can adjust reasoning effort according to task complexity to optimize latency.

In terms of architecture, they adopt a Mixture-of-Experts (MoE) approach to maximize computational efficiency. gpt-oss-120b activates 5.1B parameters per token out of a total of 117B, while gpt-oss-20b activates 3.6B out of a total of 21B. They use RoPE and Grouped Multi-Query Attention, and support context lengths of up to 128k.