Model · OpenAI · Text
gpt-oss (20b / 120b)
OpenAI's open-weight reasoning models under Apache 2.0. The headline gpt-oss-20b runs on a 16 GB machine; gpt-oss-120b fits one 80 GB data-centre GPU. Both offer low / medium / high reasoning effort, tool use and agentic coding.
- Modality
- Text
- License
- Apache 2.0 (Open)
- Parameter size
- 20b: 21B total, 3.6B active (MoE). 120b: 117B total, 5.1B active (MoE)
- Context window
- 131,072 tokens 131,072 tokens (max_position_embeddings in the published config) for both sizes.
- Released
- August 5, 2025
- Last verified
- October 28, 2026
- Runs locally
- Yes
- Also handles
- reasoning, code
What you need to run it
- Mac
- Yes. Any Apple Silicon Mac (M1 or newer) with at least 16 GB (32 GB recommended) of memory. Newer chips are faster, but memory is what decides whether it runs.
- Windows / Linux PC
- NVIDIA works best. A graphics card with at least 16 GB (24 GB recommended) of video memory.
- No graphics card
- Works. Runs on a normal computer's processor, just slower.
The 20b model fits a 16 GB Mac or graphics card per OpenAI; the 120b needs about 80 GB of graphics memory or a 96 GB or larger Mac.
Numbers are for: Official 4-bit (MXFP4) gpt-oss-20b, about 14 GB. Based on the maker’s own figures; where the maker gives only a graphics-card figure, the Mac memory is our estimate.
Strengths
- Apache 2.0 — no revenue caps, user limits or branding rules
- gpt-oss-20b runs within 16 GB of memory, so a decent laptop or a 16 GB graphics card can host it
- gpt-oss-120b fits a single 80 GB GPU (NVIDIA H100 or AMD MI300X class)
- Reasoning effort is adjustable (low, medium, high) to trade speed for depth
- Built-in function calling, web browsing, Python execution and Structured Outputs for agent work
- Full chain-of-thought is visible for debugging
Weaknesses
- Text only — no image or audio input
- Must be prompted in OpenAI's harmony response format; other prompt formats do not work correctly
- gpt-oss-120b needs a data-centre GPU, not a consumer card
- Weights are quantised (MXFP4) on the MoE layers, so check that your runner supports it
Try it
| Where | Type | Notes |
|---|---|---|
| Hugging Face (20b) | weights | Apache 2.0 |
| Hugging Face (120b) | weights | Apache 2.0 |
| Ollama | local | ollama run gpt-oss:20b |
| gpt-oss reference code | local | OpenAI's PyTorch, Triton and Apple Metal reference implementations |
Official sources
- Model card (20b) model
- Announcement announcement
- Model card paper paper
- GitHub github
Change log
- — Added hardware requirements (Mac support, memory, NVIDIA). Source: OpenAI model card / github.com/openai/gpt-oss; Ollama gpt-oss page (14 GB and 65 GB).
- — Initial entry. Added 2026-10-28 from the popularity sweep: 20.5k GitHub stars, 5.89M Hugging Face downloads/month for gpt-oss-20b and 2.86M for gpt-oss-120b. Announced by OpenAI 2025-08-05 (Hugging Face repos created 2025-08-04). Apache 2.0, 131,072-token context, free to self-host — no API price applies. Both sizes share this entry; 20b is the headline because it runs on ordinary hardware.
Esc