OpenAI's open-weight model line for self-hosting under a permissive license.
Specifications
- Provider
- OpenAI
- Type
- Open-source / open-weight
- Modality
- Text
- Category
- Language model
- Context window
- 128K
- Parameters
- Mixture-of-experts, ~120B total parameters
- License
- Apache 2.0
- Released
- August 5, 2025
What it was trained for
An open-weight large language model released by OpenAI for local and self-hosted deployment, built for reasoning and agentic tasks without depending on a hosted API.
Best for
- ▸On-premises and air-gapped deployments
- ▸Privacy-sensitive workloads
- ▸Fine-tuning and customization
- ▸Agentic and tool-using applications
- ▸Cost-controlled self-hosted inference
Capabilities
Open weightsConfigurable reasoning effortFunction calling / toolsSelf-hostable
Performance & positioning
A capable open-weight option that brings strong reasoning to self-hosted environments, competitive among openly available models while running on a single high-memory accelerator.