gpt-oss-120b & gpt-oss-20b Model Card

Summary

OpenAI's model card for gpt-oss-120b and gpt-oss-20b, its open-weight reasoning models released in August 2025. Both are mixture-of-experts transformers trained with large-scale distillation and reinforcement learning, tuned for agentic use (browsing, Python, developer functions). Weights, inference code, tool environments, and tokenizer are under Apache 2.0. The announcement page (https://openai.com/index/introducing-gpt-oss/) blocks automated fetches, so details here come from the paper's abstract and the Hugging Face model cards.

Key points

• gpt-oss-120b: 117B total, 5.1B active parameters; fits one 80 GB GPU (H100, MI300X).
• gpt-oss-20b: 21B total, 3.6B active; runs in 16 GB of memory.
• MoE weights are post-trained in MXFP4 4-bit quantisation, which is what makes those memory figures possible.
• Reasoning effort low, medium, or high, set in the system prompt ("Reasoning: high").
• Must be used with OpenAI's "harmony" response format; otherwise it does not work correctly.
• Built-in support for function calling, web browsing, Python execution, and structured outputs.
• Model cards: gpt-oss-120b, gpt-oss-20b. Context length is not stated on those pages.

Why it matters

gpt-oss was OpenAI's first open-weight language model since GPT-2, and it is still one of the few permissively licensed open models from a US frontier lab. Shipping MoE weights already in 4-bit made a strong reasoning model practical on a single GPU or laptop.