OpenAI goes open-weight with gpt-oss reasoning models

gpt-oss is OpenAI's new family of open-weight reasoning models, released under the flexible Apache 2.0 license. Marking the company's first open-weight language models since GPT-2, the release comes in two sizes: gpt-oss-120b and gpt-oss-20b.

Basically, developers can now run frontier-class reasoning and agentic workflows completely offline on their own hardware, complete with full chain-of-thought (CoT), structured outputs, and tool use like Python execution or web search.

What it does

Both models use a mixture-of-experts (MoE) architecture. While their total parameter counts sit at 117b and 21b, they only activate 5.1b and 3.6b parameters per token respectively. Both feature alternating dense and locally banded sparse attention with a native 128k context window.

Performance-wise, gpt-oss-120b achieves near-parity with OpenAI o4-mini on core reasoning benchmarks, while outperforming it on AIME math competitions and HealthBench. It also edges out o3-mini on Codeforces and agentic tool use in Tau-Bench.

The smaller gpt-oss-20b trades punches with o3-mini across common benchmarks and likewise beats it on math and health tasks. Just like OpenAI's proprietary o-series, developers can throttle reasoning effort between low, medium, and high with a single line in the system message.

How to run it

The weights are freely available on Hugging Face, pre-quantized in MXFP4 format right out of the box.

gpt-oss-20b runs within just 16 GB of memory, making it an easy fit for edge devices and consumer laptops. Microsoft has already rolled out GPU-optimized versions for Windows via ONNX Runtime, accessible in Foundry Local and VS Code. gpt-oss-120b fits comfortably on a single 80 GB GPU.

Deployment tooling is ready on day one across Ollama, llama.cpp, vLLM, and LM Studio. OpenAI also open-sourced its o200k_harmony tokenizer and harmony renderers in Python and Rust, alongside Metal and PyTorch inference templates.

Keep in mind that training was mostly text-only and English-focused around STEM and coding, so you will not get multimodal inputs here. You can grab the weights on Hugging Face or play with them directly in OpenAI's open model playground.

Why it matters

If you build local agents or handle sensitive data on-premises, this is an immediate download. Getting near-o3-mini reasoning on a 16 GB laptop changes the economics of private AI workflows overnight.

We think dropping fully capable reasoning weights under Apache 2.0 puts serious heat on proprietary mini models and rival open-weight labs alike. The catch is the strict English and text-only boundary, meaning multimodal setups still need cloud APIs for now.