Model Drop

gpt-oss-120b & gpt-oss-20b

August 5, 2025 2 min read

gpt-oss, which OpenAI released on August 5, 2025, is a pair of open-weight reasoning models, the larger gpt-oss-120b and the laptop-sized gpt-oss-20b, and OpenAI's first open-weight language models since GPT-2.

Model: gpt-oss-120b (openai/gpt-oss-120b)

Maker: OpenAI

Ship date: August 5, 2025

Pricing: Free (open weights)

Context: 128K tokens

Weights: Open (Apache 2.0)

Where you can use it:

Hugging Face weights
Ollama, LM Studio, vLLM and cloud hosts

When to reach for it:

Private reasoning on a single 80GB GPU
Skip it for images: it’s text only

How it compares:

Aimed at o4-mini
Behind o4-mini on 5 of OpenAI’s 7 tests

What makes it different:

OpenAI’s first open LLM since GPT-2
Low, medium and high reasoning effort

Early reactions: Simon Willison called the pair “very impressive” and was surprised small open models came anywhere near o4-mini and o3-mini.

Model: gpt-oss-20b (openai/gpt-oss-20b)

Maker: OpenAI

Ship date: August 5, 2025

Pricing: Free (open weights)

Context: 128K tokens

Weights: Open (Apache 2.0)

Where you can use it:

Hugging Face weights
Runs locally in 16GB of memory

When to reach for it:

On-device and laptop reasoning
Skip it for images: it’s text only

How it compares:

Aimed at o3-mini
Close to the 120b on coding

What makes it different:

Small enough for a laptop
Low, medium and high reasoning effort

Early reactions: Simon Willison called the pair “very impressive” and was surprised small open models came anywhere near o4-mini and o3-mini.

The benchmarks

OpenAI ran both models at high reasoning effort and compared them only with its own closed o3, o4-mini and o3-mini, not with other open models. Rows show gpt-oss-120b, with gpt-oss-20b listed as a rival. o3-mini took AIME without tools, so it's left out of that row; the Humanity's Last Exam and Codeforces rows compare tool-using runs only. The AA scores are for the high-effort variants.

OpenAI's numbers

GPQA Diamond (PhD-level science, no tools): 80.1%. +8.6 over gpt-oss-20b (71.5%), −1.3 under o4-mini (81.4%), +3.1 over o3-mini (77.0%), −3.2 under o3 (83.3%).
MMLU (college-level exams): 90.0%. +4.7 over gpt-oss-20b (85.3%), −3.0 under o4-mini (93.0%), +3.0 over o3-mini (87.0%), −3.4 under o3 (93.4%).
Humanity's Last Exam (with tools) (expert questions): 19.0%. +1.7 over gpt-oss-20b (17.3%), +1.3 over o4-mini (17.7%), −5.9 under o3 (24.9%).
AIME 2025 (math contest, with tools): 97.9%. −0.8 under gpt-oss-20b (98.7%), −1.6 under o4-mini (99.5%), −0.5 under o3 (98.4%).
SWE-bench Verified (real GitHub fixes): 62.4%. +1.7 over gpt-oss-20b (60.7%), −5.7 under o4-mini (68.1%), +13.1 over o3-mini (49.3%), −6.7 under o3 (69.1%).
TAU-bench Retail (function calling): 67.8%. +13.0 over gpt-oss-20b (54.8%), +2.2 over o4-mini (65.6%), +10.2 over o3-mini (57.6%), −2.6 under o3 (70.4%).
Codeforces (competition coding Elo, with tools): 2622. +106 over gpt-oss-20b (2516), −97 under o4-mini (2719), −84 under o3 (2706).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 11.6. +2.6 over gpt-oss-20b (9.0), −5.1 under o4-mini (16.7), −1.5 under DeepSeek R1 0528 (13.1), −1.1 under Qwen3-235B-A22B-Thinking-2507 (12.7).