gpt-oss-120b

gpt-oss-120b is an open-weight reasoning model from OpenAI, released on 5 August 2025 alongside the smaller gpt-oss-20b. The models are provided under the Apache 2.0 license with full access to the weights, permitting commercial and non-commercial deployment as well as fine-tuning.

The model uses a transformer-based mixture-of-experts (MoE) architecture: 117B total parameters, of which 5.1B are active per token; 128 experts (4 active per token) and 36 layers. Thanks to MXFP4 quantization of the MoE weights, the model fits on a single 80GB GPU (e.g. NVIDIA H100 or AMD MI300X).

gpt-oss-120b is a reasoning model with configurable effort levels (low/medium/high) and a visible chain of thought. It supports a 131,072-token context window, tool use (web browsing, Python code execution), function calling and structured outputs; it requires the harmony response format. It accepts and produces text only. Its knowledge cutoff is June 2024.

In evaluations (high reasoning mode) the model scores 96.6% on AIME 2024 and 97.9% on AIME 2025 (with tools), 80.9% on GPQA Diamond, 90.0% on MMLU, 2622 Elo on Codeforces and 57.6% on HealthBench. OpenAI reports that gpt-oss-120b achieves results comparable to the o4-mini model. The model is available for download on Hugging Face and through many inference providers.