gpt-oss

OpenAI's open-weight release of August 2025, its first language-model weights since GPT-2, and a backfill here: two MoE reasoning models under Apache 2.0. gpt-oss-120b has 117B total and 5.1B active parameters (36 layers, 128 experts, 4 active per token) and is sized for a single 80 GB GPU; gpt-oss-20b has 21B total and 3.6B active (24 layers, 32 experts) for local and low-latency use. Both share a 201,088-token vocabulary, alternate 128-token sliding-window and full attention, run a 131,072-token context, and expose a configurable reasoning effort (low, medium, high). They are trained on and must be served with OpenAI's harmony response format, and support tool use and chain-of-thought output. The HuggingFace repositories have passed five and six million downloads respectively.

On Artificial Analysis the 120B at high effort scores 12 on Intelligence Index v4.3 and the 20B scores 9; both were near the top of the open-weights table at release but sit in the lower half after the 2026 recalibrations.