What Are Open-Weight Models?
Open-weight models are models whose weights, architecture and inference code are publicly released, so organisations can run, adapt and fine-tune them on their own infrastructure. They are not always open source in the sense of the Open Source Initiative, which also asks for information about training data and the training code. The Thoughtworks Technology Radar Vol 35 places open-weight models in Trial, reporting good results from the DeepSeek, GLM and Kimi families and pointing to lower total cost of ownership, deployment flexibility, less vendor lock-in and stronger data sovereignty. Self-hosting takes more operational effort, but major cloud providers now offer managed open-weight models. This page summarises the Radar view in my own words and lists definitions, access routes, fine-tuning tools, evaluation sources and policy material.
The Radar View
4 entriesThe Thoughtworks Technology Radar Vol 35 (November 2026) places open-weight models in Trial and lists blanket frontier model usage as a practice to be cautious about.
Definitions, Openness & Licences
6 entriesWhat open weights do and do not give you, and how openness is assessed.
Model Families Named by the Radar
3 entriesHugging Face organisations for the three families Thoughtworks evaluated. See the Open-Source LLMs directory on this site for current releases.
Managed & Hosted Access
9 entriesThe Radar notes that major cloud providers now offer managed open-weight models, which lowers the barrier to adoption.
Self-Hosting & Local Inference
4 entriesEngines for running open-weight models on your own infrastructure, which the Radar says takes more operational effort.
Fine-Tuning Open Models
4 entriesLibraries for private fine-tuning on domain data.
Evaluation & Comparison
2 entriesIndependent data for comparing open-weight and proprietary models. Test on your own tasks before choosing.
Policy & Risk
4 entriesGovernment and standards material on models whose weights are widely available.
Related Pages
5 entriesRelated directories on this site.
Frequently Asked Questions
What are open-weight models?
Open-weight models are models whose weights, architecture and inference code are publicly released, so you can download and run them yourself. Thoughtworks uses this definition in its Technology Radar.
Are open-weight models the same as open-source AI?
No. The Open Source Initiative's Open Source AI Definition asks for more than weights: information about the training data, the complete code used to train and run the system, and parameters, all under open terms. Many open-weight releases do not meet that bar, and their licences can add restrictions.
Why does Thoughtworks put open-weight models in Trial?
It reports good results from DeepSeek, GLM and Kimi models on production work needing complex reasoning and good code generation. It cites lower total cost of ownership, deployment flexibility, less vendor lock-in and stronger data sovereignty, and notes that self-hosting takes more operational effort.
How should a team start with open-weight models?
Try a managed service from a cloud provider first, test candidates on your own tasks, and route routine work to smaller or open models. Thoughtworks lists blanket frontier model usage, sending every task to the most capable model, as a practice to be cautious about.
What are the risks of open-weight models?
Licence terms can restrict use, self-hosting adds operations and security work, and releases vary in how well they are documented and tested. Policy bodies such as the US NTIA and the European Commission have published material on models with widely available weights.