gpt-oss Review

gpt-oss is the pair of open weight reasoning models OpenAI released on August 5, 2025, its first open language models since GPT-2. The larger gpt-oss-120b is a mixture of experts with 117 billion parameters, 5.1 billion of them active per token, and fits on a single 80GB GPU; gpt-oss-20b runs in 16GB of memory, enough for a well specified laptop. Both carry a 131,072 token context window, adjustable low, medium and high reasoning effort, built in tool use and the permissive Apache 2.0 licence.

At launch the consensus called them the best open models for their size, close to o4-mini and o3-mini on reasoning and maths, with competent code. Independent testing soon placed them behind the largest Chinese open models on raw capability, OpenAI's own evaluations show them hallucinating more than o4-mini on factual questions, and their knowledge stops in June 2024. Long agentic coding runs are where the gap to larger models shows first.

A year on they remain a sensible default for private and on premise work: free, cheap to host, quick thanks to the small active parameter count, and easy to run through Ollama, LM Studio or the major clouds. For frontier coding or research depth, newer open models such as DeepSeek V4, Qwen3.5 and Gemma 4 have moved past them.