open-weight model
A machine learning model whose weights (learned parameters) are publicly accessible, enabling local deployment, auditing, and customization without proprietary restrictions. Differs from closed-weight models (e.g., most commercial LLMs) where weights are withheld.
Key characteristics:
- Weights available for download (e.g., via public repositories)
- May not include full training code/data (only weights)
- Enables offline use, customization, and transparency
- Often distributed under permissive licenses
Recent examples:
- OpenAI’s
[[concepts/gpt-oss-20b|gpt-oss-20b]]and[[concepts/gpt-oss-20b|gpt-oss-20b]] - Mistral AI’s Mistral Large 4 (“Le Chonk”), a 1-trillion parameter, open-weight, open-source model developed in Europe. It utilizes a Mixture of Experts (MoE) architecture with 49 billion active parameters and supports multimodal inputs. See European AI Sovereignty: Mistral Large 4 Capabilities, Performance, Challenges for detailed analysis.
References: