OpenAI gpt-oss

Ollama announced on August 5, 2025 that it has partnered with OpenAI to bring OpenAI's gpt-oss open weight models to Ollama and its community. The release covers two models, 20B and 120B parameters, described as designed for powerful reasoning, agentic tasks, and versatile developer use cases.

According to Ollama, the models offer agentic capabilities including native function calling, optional built-in web search provided by Ollama, python tool calls, and structured outputs. They provide full chain-of-thought access for debugging and trust in outputs, configurable reasoning effort across low, medium, and high settings, fine-tuning through parameter customization, and release under a permissive Apache 2.0 license.

The gpt-oss-20b model is designed for lower latency, local or specialized use cases, while gpt-oss-120b is designed for production, general purpose, high reasoning use cases. OpenAI applies quantization to reduce memory footprint, post-training the mixture-of-experts weights to MXFP4 format at 4.25 bits per parameter. Since the MoE weights account for 90+% of the total parameter count, quantizing them to MXFP4 lets the smaller model run on systems with as little as 16GB of memory and the larger model fit on a single 80GB GPU.

Ollama said it supports the MXFP4 format natively without additional quantizations or conversions, with new kernels developed for its new engine. Ollama also stated it collaborated with OpenAI to benchmark against OpenAI's reference implementations to ensure the same quality. NVIDIA and Ollama are advancing their partnership to boost model performance on NVIDIA GeForce RTX and RTX PRO GPUs, which Ollama said enables users on RTX-powered PCs to leverage gpt-oss capabilities. Ollama said it will publish an in-depth engineering post on the model in the future. The models can be run via ollama run gpt-oss:20b and ollama run gpt-oss:120b.