Open-weight versus closed AI models: what you actually get, and what to choose

Every AI model you can use falls into one of two broad groups. With a closed model you send requests to the vendor's servers and pay per use. With an open-weight model, the vendor publishes the trained model file, and you can run it on hardware you control. The choice affects cost, data handling and how much of the model you can change.

The vocabulary

Weights are the numbers a model learned in training. Publishing them lets anyone run the model. Open source, in the Open Source Initiative's definition, asks for more: freedom to use, study, modify and share the system, with the preferred form for modification including detailed information about training data, the complete training and inference code, and the parameters ^[1]. Many "open" models release weights and some tooling but not the training data, so "open-weight" is the more accurate label.

Closed models, such as Anthropic's Claude Opus 4.6 (priced at $5 and $25 per million input and output tokens up to 200,000-token prompts), are reached only through an API or product ^[4].

Two recent open releases

OpenAI's gpt-oss-120b is released under the permissive Apache 2.0 license. It has 117 billion parameters in total, of which 5.1 billion are active for any token, and it is designed to fit on a single 80 GB GPU such as an Nvidia H100 ^[2]. The model card lists support for function calling, web browsing and Python execution, and describes fine-tuning on a single H100 node ^[2].

In February 2026, Zhipu AI released GLM-5 under the MIT license. Its model card describes a mixture-of-experts model with 744 billion parameters (40 billion active), links the asynchronous reinforcement-learning infrastructure used for training, and lists several serving frameworks and accelerator deployment guides ^[3].

How close are they?

On SWE-bench Verified, a software-engineering test, the model card lists GLM-5 at 77.8% against 80.9% for Claude Opus 4.5 ^[3]. These are the vendor's own reported scores and have not been independently verified here; benchmark scores do not necessarily reflect real-world performance.

Choosing between them

The hardware requirement is the hidden cost: a 117-billion-parameter model that fits on one 80 GB GPU is practical, while a 744-billion-parameter model is a multi-GPU deployment ^[2][3].

A practical rule

Start with a closed model to prove the use case, because it is the fastest route to a working result. Move a workload to an open-weight model when you can name a reason: data that cannot leave your network, a steady volume that makes owned hardware cheaper, or a need to modify the model itself.

For businesses: pick the model per workload, not per company, and record the exact release and licence name (for example Apache 2.0 or MIT) of any open-weight model you deploy.