Particle.news
Download on the App Store
118 ARTICLES
·
last mo.

OpenAI Releases GPT-OSS 120B and 20B Under Apache 2.0 License

Released under Apache 2.0, the text-only GPT-OSS models run on hardware from laptops to single GPUs, having passed rigorous safety evaluations to meet growing demands for transparent AI.

Image
PCs powered by an RTX 5090 GPU can get performance of up to 256 tokens per second when running OpenAI's new models.
Image
OpenAI logo is seen in this illustration taken May 20, 2024. REUTERS/Dado Ruvic/Illustration/File Photo

Overview

  • OpenAI published gpt-oss-120b (117 billion parameters) and gpt-oss-20b (21 billion parameters) weights on platforms like Hugging Face under an Apache 2.0 license, permitting free commercial use, modification and redistribution.
  • The 117 billion-parameter gpt-oss-120b runs on a single Nvidia GPU; the 21 billion-parameter gpt-oss-20b requires just 16 GB of memory, enabling deployment on consumer laptops.
  • Benchmark tests show gpt-oss-120b performing on par with OpenAI’s o4-mini and both models outpacing DeepSeek’s R1 on Codeforces and academic assessments.
  • OpenAI’s safety regime—including adversarial fine-tuning, third-party reviews and high-risk filtering—found no evidence that the open models exceed internal danger thresholds for cyber or biological misuse.
  • The move responds to intensifying competition from Chinese AI labs and U.S. policy calls for open AI, marking a strategic pivot from OpenAI’s closed-source API business model.