An open-weight model has trained parameters, or weights, available to download and run. That makes self-hosting possible, but it does not automatically make the model open-source or grant unrestricted use. Open-source AI also requires access to the components and information needed to study, modify, and reproduce the system. section 1.1.6 section 1.4
For a team considering self-hosting, the decision is not simply whether an open-weight model is good enough. It is which model fits the workload, what infrastructure it needs, and whether the organization can operate it reliably. Self-hosting can support privacy, cost, control, vendor-risk, or regulatory goals, but may not fit low-volume workloads or teams without platform ownership. section 1.1.6 section 1.8
A practical shortlist starts with these checks:
- Review the license before testing quality. Custom licenses can restrict commercial use, training on outputs, or other activities; terms can differ within a family or change between generations. section 1.4 section 1.8
- Check the upstream lineage of derivatives. A fine-tune incorporates base-model weights, while distillation can raise different questions about source-model terms and training data. section 1.4
- Match the model and serving setup to the workload. Consider memory needs and load time, as well as what happens when a machine fails or responses slow down. section 1.1.6
- Use popularity as a starting point, not proof of production fit. The landscape includes fine-tunes and quantized versions of original models, alongside many other derivatives. section 1.3
Table 1.2 License orientation
|
License category
|
Commercial use
|
Modification
|
Redistribution
|
Practical meaning
|
Example families
|
|
Permissively licensed weights, for example under Apache
2.0 or MIT
|
Permitted
|
Permitted
|
Permitted
|
Generally suitable for production, provided that the
license conditions are followed.
|
Many Qwen and Mistral models, gpt-oss, Muse Glimmer
|
|
Open-weight with a custom license
|
Usually permitted
|
Depends on the license
|
Depends on the license
|
Depends on the license
|
Llama, Gemma, Kimi K3
|
|
Research or non-commercial
|
Usually not permitted
|
Limited
|
Limited
|
Not suitable for commercial production use.
|
Some research releases, such as models under the Mistral
Research License
|
Compare permissive, custom, and non-commercial license categories before reviewing the exact terms for a candidate model and its upstream lineage. section 1.4
Choosing for a law firm
Meridian's IT team begins with an open-model search, then faces a landscape of original families and derivatives. Because license terms matter to the firm, a candidate's permissions and upstream lineage need review before it becomes a production choice. section 1.3 section 1.4
Serving tools also differ by deployment size. llama.cpp and tools built on it, including Ollama and LM Studio, are popular for a single machine; vLLM serves many users at once. Kubernetes can manage vLLM across multiple machines when one is no longer enough. section 1.1.6


