OllamaFrom the builder
OpenAI gpt-oss
Ollama partnered with OpenAI to offer the open-weight gpt-oss models in 20B and 120B sizes, built for reasoning, agentic tasks and varied developer uses under the Apache 2.0 license. Their mixture-of-experts weights are quantized to MXFP4 at 4.25 bits per parameter, letting the 20B model run on systems with 16GB of memory and the 120B model fit on a single 80GB GPU. Ollama supports MXFP4 natively and benchmarked against OpenAI's reference implementations to match quality.
Ollama is supporting the MXFP4 format natively without additional quantizations or conversions.
Saved . A shortened excerpt from the official blog; the original publisher retains authorship.