AI decisions Architecture and technology
Open-source vs proprietary LLMs: how should an enterprise choose?
Choose by use case, not by camp. Proprietary models via API are usually the fastest way to strong general capability with little operational burden; open-weight models win when you need full control of data and deployment, deep customisation or predictable cost at high volume, and you have the team to run them. Most enterprises end up with both, behind an abstraction layer that lets them switch.
The options
Open-source / open-weight LLMs
Models whose weights you can download and run on your own infrastructure or a cloud you control, under the terms of their licence.
Proprietary LLMs
Models you access as a service through the provider's API or a cloud marketplace, without access to the weights.
Side by side
| Criterion | Open-source / open-weight LLMs | Proprietary LLMs |
|---|---|---|
| Control over data and deployment | Full: you decide where inference runs and what is logged. | Bound by the provider's terms, regions and data-handling options. |
| Capability on hard, general tasks | Narrowing gap, but uneven from task to task; test on your own cases. | Usually the strongest on complex reasoning at any given moment. |
| Customisation | Deep: fine-tuning, distillation, quantisation, own serving stack. | Limited to what the provider exposes (prompting, some fine-tuning options). |
| Operational burden | Yours: infrastructure, scaling, patching, monitoring, security. | Mostly the provider's; you manage integration and usage. |
| Cost profile | Fixed infrastructure and talent; attractive at high, steady volume. | Pay per use; attractive for variable or low volume, harder to predict at scale. |
| Change control | You choose when to upgrade; the model does not change under you. | Versions are updated or retired on the provider's schedule. |
| Licensing | Varies widely: some licences restrict use, scale or sectors. Read them. | Commercial contract; check data use, liability and exit terms. |
| EU AI Act position | If you substantially modify or fine-tune, you may take on provider duties. | The model provider carries GPAI obligations; you remain responsible as deployer of your system. |
Choose Open-source / open-weight LLMs when…
- Data cannot leave infrastructure you control, by regulation, contract or policy.
- The task is narrow, high-volume and well defined (classification, extraction, routing), where a tuned smaller model performs well.
- You need a model version that stays fixed for audit or validation reasons.
- You have, or can build, the MLOps and security capability to run models in production.
Choose Proprietary LLMs when…
- You need the strongest available capability for complex reasoning or generation.
- Volume is uncertain or low, and paying per use is cheaper than running infrastructure.
- Speed matters more than control, and the data involved can be processed under the provider's terms.
- Your team is small and better spent on the use case than on operating models.
When to combine them
The practical answer is a portfolio. Route complex, low-volume reasoning to proprietary models and high-volume, well-bounded or sensitive tasks to open-weight models you host. The condition is an abstraction layer between applications and models, plus your own evaluation set, so each switch is a configuration decision backed by evidence rather than a rewrite.
Common mistakes
- Treating open source as free: hardware, talent, security and monitoring are the real cost.
- Choosing from public leaderboards instead of testing on your own data and edge cases.
- Assuming “open” means unrestricted; some licences limit commercial use or scale.
- Coding applications directly against one provider's API, which turns a model choice into lock-in.
- Forgetting that fine-tuning or rebranding a model can change your obligations under the EU AI Act.
How Thinkia approaches it
We do not pick sides. We help clients build the capability to choose: an evaluation set from their own cases, clear criteria per use case (data sensitivity, volume, latency, required capability, change control) and a total cost view that includes infrastructure and people, not only tokens.
Architecturally, we put a model-agnostic layer between applications and models. Synapse does this: its gateway enforces corporate SSO and usage limits, and its routing can send confidential workloads to local models before commercial endpoints. Switching the model does not mean touching the front end or the business logic. We work with proprietary providers and with open models through Hugging Face, and the choice follows the use case.
On regulation, we map who is provider and who is deployer for each system. Under Regulation (EU) 2024/1689, obligations for general-purpose AI model providers apply from 2 August 2025, with a partial exemption for models released under free and open-source licences that does not cover models with systemic risk. If a client fine-tunes or white-labels a model, we review whether that shifts provider duties to them. This is orientation, not legal advice.
Thinkia products involved
- SynapseGoverned agentic platform: agents, models, costs and data in one place.
- EU AI Act governance guideRisk tiers, timeline, roles and a 20-point checklist. Not legal advice.
Related AI solutions
Frequently asked questions
Are open-source models good enough for enterprise use now?
For many tasks, yes, especially narrow ones that can be tuned. Their performance is often uneven across similar tasks, so the only reliable answer comes from testing on your own data, including edge cases and adversarial inputs.
Is open source cheaper than paying for an API?
It depends on volume and team. At steady high volume, self-hosting can be cheaper per request; at low or variable volume, the fixed cost of infrastructure and specialists usually makes the API cheaper. Compare total cost over the life of the use case, not price per token.
Is open-weight the same as open source?
Not quite. Many models publish their weights but not their training data or code, and their licences range from very permissive to restrictive. Read the licence before building on a model, especially for commercial or large-scale use.
Does the EU AI Act favour open-source models?
It gives free and open-source general-purpose models a partial exemption from some provider documentation duties, but not when the model has systemic risk. Your obligations as deployer depend on the use case, whatever the model. Check the text on EUR-Lex; this is not legal advice.
How do we avoid being locked in either way?
Keep applications independent of any single model API through a gateway or abstraction layer, keep your prompts, evaluation sets and data in your own hands, and negotiate exit terms with providers. Then the model becomes a replaceable component.
Keep exploring
Related decisions
- Sovereign or on-prem AI vs cloud AI APIs: where should your models run?
- Single AI vendor vs multi-model strategy: should you bet on one provider?
- RAG vs fine-tuning: which one does your enterprise use case need?
- EU AI Act provider vs deployer: which role are you, and what does each one owe?
Sectors where this decision comes up
Key terms
Thinkia articles
- Hybrid AI Strategy: Why Open-Source Models Are Now Essential
- Open-Source AI Models: A CIO's Guide to the Jagged Frontier
- Open-Source AI Regulation: Mapping the Scenarios for Enterprise Leaders
- Beyond Giants: Why a Portfolio of Specialized Models Is Your Next AI Advantage
- Frontier Model Compliance: The EU AI Act's Hidden Liability