Open-weight models provide access to the trained model parameters, allowing organizations to run and customize the AI locally, but differ from traditional open-source software by not necessarily including the original training code or datasets.
Architecture designed for enterprise efficiency
The models leverage a mixture-of-experts (MoE) architecture to optimize computational efficiency. The gpt-oss-120b activates 5.1 billion parameters per token from its 117 billion total parameters, while gpt-oss-20b activates 3.6 billion from its 21 billion parameter base. Both support 128,000-token context windows and are released under the Apache 2.0 license, enabling unrestricted commercial use and customization.
The models are available for download on Hugging Face and come natively quantized in MXFP4 format, according to the statement. The company has partnered with deployment platforms, including Azure, AWS, Hugging Face, vLLM, Ollama, Fireworks, Together AI, Databricks, and Vercel to ensure broad accessibility.
Read the full article here
