gpt-oss:120b-cloud

Feature highlights — shared with gpt-oss-20b

• Agentic capabilities — native function calling, web browsing, Python tool calls, and structured outputs.
• Full chain-of-thought — complete access to the model's reasoning process for easier debugging and increased trust.
• Configurable reasoning effort — adjust reasoning effort (low, medium, high) based on the use case and latency needs.
• Fine-tunable — fully customize through parameter fine-tuning.
• Apache 2.0 license — permissive licensing for experimentation, customization, and commercial deployment.

Quantization

MoE weights are quantized to MXFP4 (4.25 bits per parameter), enabling the 120B model to fit on a single 80GB GPU. Ollama collaborated with OpenAI to benchmark against their reference implementations to ensure the same quality.

Best practices

• Frontier open reasoning — the larger variant for demanding reasoning and agentic tasks.
• Reasoning effort — tune low/medium/high per request to balance depth and latency.
• Deployment — single 80GB GPU locally, or served via :120b-cloud on Ollama's infrastructure.