Open weights or open source? Why the licence matters when you train on your own data

Enterprises are moving away from multi-tenant APIs to locally hosted model weights. This shift is often described as a transition to Open Source, but that terminology is frequently inaccurate. For a CTO or Legal Counsel, the distinction between Open Weights and true Open Source transcends semantic nuance: it represents a fundamental difference in how much control the business retains over its automated processes.

When a company invests in post-training a model on private, high-value data, the resulting weights become a primary business asset. If that asset is governed by a restrictive or revocable licence, the business has introduced a systemic risk into its core infrastructure.

The definition of control

The industry often conflates any model that can be downloaded with "Open Source." This is technically incorrect. Most prominent models (including Llama 3 and Mistral) are Open Weights models. They provide the model parameters but keep the training data, recipes, and full rights proprietary.

A true Open Source model, such as those released under the Apache 2.0 licence, offers unconditional freedom to use, modify, and distribute. Open Weights models, conversely, are typically governed by bespoke licences that include "Acceptable Use Policies" (AUPs). These policies can be updated at the discretion of the provider, creating a "catastrophic forgetting" of legal rights for the licensee.

Feature	Open Source (e.g., Apache 2.0)	Open Weights (e.g., Llama/Mistral)
Commercial Use	Unrestricted	Often restricted by revenue or user count
Modification	Fully Permitted	Permitted under specific AUP constraints
Revocation	Irrevocable	Can be revoked for AUP violations
Transparency	Full (Data + Training Code)	Partial (Weights Only)
Sovereignty	Absolute	Conditional

Scrutinizing the Acceptable Use Policy

When you fine-tune an adapter (like a LoRA) or perform a full-parameter update on your proprietary data, you are baking your institutional knowledge into the model. This specialist adapter is what allows the model to perform a business process with high precision.

Legal teams must evaluate if the upstream model licence allows for this "derivative" work to be owned exclusively by the customer. Some AUPs contain clauses that prohibit using the model to improve other models, which can be interpreted broadly enough to include fine-tuning for certain competitive applications.

Weights as a business asset

Treating model weights as a temporary service is a strategic error. In a production environment, the weights are the digital distillation of your policy and expertise. Owning the specialist adapter is the only way to ensure long-term sovereign control of a business process.

If you rely on a model that can be legally "recalled" by a provider through an updated AUP, you are building on someone else's land. True sovereignty requires that the weights you run today remain yours to run ten years from now, regardless of the provider's corporate roadmap or market shifts.

The Agentsia approach: Legal and security boundaries

At Agentsia, we prioritize models that fit within the customer's legal and security boundary. We promote a "sovereign-first" architecture where the most critical business logic is handled by models with the most permissive licences.

1. Licence auditing: We analyze the base model licence to ensure it supports the customer's long-term commercial goals.
2. Adapter ownership: We ensure that all post-training artifacts created on the Modelsmith platform are the exclusive property of the customer.
3. Sovereign deployment: We favor Apache 2.0 or similarly permissive base models for core workflows where revocation would be catastrophic.

Long-term stability in AI depends on the durability of the legal right to use the weights as much as it depends on their technical quality.

If you are looking to build a fleet of specialist models that you actually own, explore our Modelsmith platform or contact our legal-engineering team.