What Is GPT-OSS? Inside OpenAI’s New Open‑Source Large Language Models

Model Overview

According to OpenAI’s announcement, the GPT‑OSS series includes two variants: GPT‑OSS‑120B with about 12 billion parameters designed for high‑throughput production inference, comparable to OpenAI’s o4‑mini and able to run efficiently on a single 80 GB GPU; and GPT‑OSS‑20B with about 2 billion parameters optimized for low latency, able to run on consumer‑grade hardware with 16 GB memory, comparable to o3‑mini.

The models use a Mixture‑of‑Experts architecture and a 4‑bit quantization scheme (MXFP4) that keeps resource usage low while delivering fast inference.

Benchmark results show strong performance on reasoning tasks, especially chain‑of‑thought and tool use. GPT‑OSS‑120B approaches o4‑mini on core reasoning benchmarks, while GPT‑OSS‑20B is suitable for edge devices and rapid prototyping. Users can select configurable inference effort levels (low, medium, high) to balance latency and quality.

A notable feature is the permissive Apache 2.0 license, allowing broad modification and commercial use without patent concerns.

OpenAI’s evaluation only compares GPT‑OSS against its own models; users can compare the reported performance (e.g., GPT‑OSS‑120B ≈ o4‑mini, GPT‑OSS‑20B ≈ o3‑mini) with other vendors’ models to assess suitability.

References

https://openai.com/index/introducing-gpt-oss/

https://huggingface.co/openai/gpt-oss-120b