OpenAI launches gpt-oss open-weight language models

OpenAI has released its first open-weight language models in over five years, marking a return to greater transparency after the proprietary GPT-3 and GPT-4 series. The new models, gpt-oss-120b and gpt-oss-20b, were introduced on 5 August 2025 under the Apache 2.0 licence, which allows for unrestricted commercial and research use.

These models are not fully open-source—OpenAI has released the model weights but not the training data or complete codebase. However, they are considered open-weight, making them accessible for fine-tuning and offline deployment.

Model architecture and technical details

The models use a Mixture of Experts (MoE) architecture, which activates a subset of neural pathways per token. The larger model, gpt-oss-120b, includes approximately 117 billion parameters, with 128 experts, 4 of which are active per token. The smaller variant, gpt-oss-20b, includes 21 billion parameters and 32 experts.

Both models support a 128,000-token context window, enabling extended text understanding and generation. They incorporate grouped multi-query attention and Rotary Positional Embeddings (RoPE), enhancing reasoning capabilities and memory efficiency.

OpenAI reports that gpt-oss-120b performs at or above the level of its proprietary o4-mini model, while gpt-oss-20b rivals o3-mini performance despite a much smaller footprint. The smaller model is optimised for consumer-grade systems with at least 16 GB of RAM.

Safety evaluation and release rationale

OpenAI conducted both internal red-teaming and third-party evaluations to assess potential misuse. The models were tested against scenarios such as code generation misuse, misinformation, and disinformation. The outcome, according to OpenAI’s preparedness framework, indicates a “low risk” of misuse at scale.

Licensing and community access

The Apache 2.0 licence allows users to download, modify, and redistribute the models with minimal restrictions. Commercial use is permitted, provided users comply with the licensing terms and safety guidelines outlined by OpenAI.

Documentation and evaluation benchmarks are available to aid adoption and responsible usage. The release does not affect OpenAI’s flagship products, such as ChatGPT, which remain closed-weight and proprietary.