GPT-OSS (OpenAI)

tl;dr: OpenAI’s first open-weight flagship in years, two sizes, Apache 2.0, MXFP4 quantization so the 20B fits in 16GB and the 120B fits on one 80GB GPU.

What it is

Two open-weight text models from OpenAI, hosted natively in Ollama’s new engine. Both are Mixture-of-Experts with the MoE weights post-trained in MXFP4 (4.25 bits/param). The full chain-of-thought is exposed in the output, and reasoning effort is configurable (low / medium / high).

• gpt-oss:20b — lower latency, local-or-specialized. 128K context.
• gpt-oss:120b — frontier reasoning. 128K context.
• Cloud-tagged variants (gpt-oss:20b-cloud, gpt-oss:120b-cloud) route to Ollama-hosted inference and burn Pro plan usage.

Pull

ollama run gpt-oss:20b
# or:
ollama run gpt-oss:120b

What you get

• Native function calling, structured outputs, and Python tool calls.
• Optional built-in web search (Ollama-side, can be enabled).
• Configurable reasoning effort — pick low/medium/high per request.
• Fine-tunable per OpenAI’s published model card.
• Apache 2.0: no copyleft, no patent rider, commercial deployment OK.

Hardware (per Ollama + OpenAI)

Gotchas

• Cloud variants need an Ollama account and Pro plan usage credits. Local run is always free.
• Thinking tokens are visible in the output — disable if you don’t want them in logs or downstream prompts.
• Web search is Ollama-side and opt-in; it is not the same as OpenAI’s responses web search tooling.

Related: Ollama for the runtime, llama.cpp for the engine, Open WebUI for a chat UI, LM Studio for a GUI runner.

Untested here — facts from ollama.com/library/gpt-oss and openai.com fetched 2026-09-10.