# Open-Source Models or Commercial Models

Author: Marin Niehues, Niehues Consulting · Published: 2025-03-31 · Updated: 2026-11-16 · AI Adoption & Operations
Source: https://www.niehuesconsulting.com/blog/open-source-vs-commercial-ai-models/

Whether open-source or commercial models fit better is decided by the use case, and most companies are best served by a portfolio of both. The question is often treated as a matter of principle, although it arises afresh for every application. This article sets out both options, names five criteria for the choice and explains what the license, the hardware and the AI Act mean for open models. This article provides an overview and does not constitute legal advice.

**Model choice is a delivery decision. What matters is which option can be operated and paid for over the long term in a specific process.**

## What distinguishes open-source models from commercial models?

Commercial models are operated by a provider who makes them available through an interface, while open models publish their weights and are run by the company itself or by a service provider. With a commercial model, the provider is responsible for infrastructure and updates, and the company pays by usage or license. With an open model, the company decides on location and version as well as on adaptation, including further training with its own data.

For precision: the Open Source Initiative published a definition of open-source AI in 2024. Under it, open source in the strict sense requires the weights, the code and sufficient information about the training data to be available under terms that permit free use, the modification and the redistribution. Most models marketed as open do not fully meet this definition. The technically correct term for them is open weights. In business usage, open source has nevertheless remained the umbrella term, and this article follows it.

The two types form a spectrum. Between them sit services that run open models and bill by usage, and commercial providers that offer their models in a European region or in a dedicated environment. For each use case, the question is therefore where on this spectrum it belongs. The article [Large language models compared](/blog/large-language-models-comparison/) gives a market overview.

**Key terms**
Open source in the strict sense: weights, the code and sufficient information about the training data are available under a license that permits free use, the modification and the redistribution, as the Open Source Initiative definition requires.

Open weights: the model can be downloaded and run independently, but the license may contain restrictions.

Commercial model: access is through a provider on its terms, and operation sits with the provider.

## Which open-source language models are there?

The best-known open model families include Llama from Meta, the models of the French provider Mistral AI, Qwen from Alibaba, DeepSeek from China and Gemma from Google. Each family spans several sizes, from small models for a single graphics card to very large models for data centers. The list implies no ranking; what counts is performance in your own process.

The families differ above all in their licenses. Some Mistral and Qwen models are released under the Apache 2.0 license, which permits commercial use and modification with few obligations. DeepSeek has released models under the MIT license. Llama and Gemma follow their manufacturers' own license terms with usage policies. For models from third countries, one distinction applies: running the weights yourself keeps the data in your own environment, whereas using the manufacturer's online services transfers the data to its infrastructure.

## Can Llama be used commercially?

Yes, Meta's Llama license permits commercial use, but it contains conditions that go beyond a classic open-source license. Companies with more than 700 million monthly active users need a separate license from Meta. In addition, there is a usage policy listing prohibited purposes, and attribution obligations apply on redistribution. Individual releases contain further restrictions, so the license of the specific release in use is what applies.

One rule follows for practice: the license review is a fixed step in every model selection, and it comes before the prototype. Its legal assessment is a matter for qualified legal counsel. Leaving it until just before rollout risks having to rebuild a working prototype.

## When does an open model pay off?

An open model pays off with high data protection requirements, clearly defined tasks and high, steady volume, provided the company secures operation itself or through a service provider. A sound decision results when every use case is assessed against the same five criteria.

First, the protection requirements of the data. The more sensitive the data, the greater the value of an environment the company controls itself. Running open models in-company keeps the data within its own infrastructure. Commercial providers also offer contractually secured environments. Which set-up is sufficient for which data class is settled by the data protection review.

Second, quality in the specific use case. General leaderboards say little about performance in your process. What matters is an evaluation with your own examples from the business function: representative cases with defined success criteria, run with both model types under identical conditions. For clearly defined tasks such as classification or extraction, smaller open models are often sufficient. For open and demanding tasks, the strongest commercial models lead the performance frontier, but the gap to the best open models has become small; your own evaluation decides.

Third, operational capability. Running a model yourself requires capacity for deployment and monitoring as well as for updates and security. These capabilities can be built or bought, and they must be backed by assigned responsibilities before implementation starts. A model without a named operator does not reach regular operation.

Fourth, the cost structure. Commercial models incur variable costs per use that rise with volume. Open models in self-operation incur mainly fixed costs for hardware and staff, which fall per request when volume is high and steady; when run by a service provider, they too are billed by usage. The break-even point differs for each use case, and the calculation belongs in the business case. With low or fluctuating volume, the cost logic favors the variable option.

Fifth, the ability to switch. Providers change prices and model versions, and a model that fits today will be retired at some point. The ability to switch comes from a layer between application and model, from an evaluation set that makes a replacement testable in a short time, and from documented instructions to the model.

**The model types at a glance**
Criterion | Open-source models | Commercial models | With Niehues Consulting
Data control | Full control in self-operation | Contractually secured, terms set by the provider | Data classes assigned per use case, suitable environment chosen
Quality | Strong for clearly defined tasks | Performance frontier for open, demanding tasks | Decision based on the business function's own cases
Operating effort | High in self-operation, own capabilities required | Low, the provider operates | Operating model agreed before the start, internal or external
Cost profile | Fixed in self-operation, falls per request with volume | Variable, rises with volume | Business case per application with a volume path
License | Check per model, ranging from Apache 2.0 to manufacturer license | Provider's terms of use | License review before the prototype
Ability to switch | High with standardized formats | Set by the provider | Model layer and evaluation set from the outset

## What hardware do you need to run a language model yourself?

To run a model yourself, you mainly need graphics cards with enough memory for the model weights, and the requirement is calculated from the number of parameters. The calculation provides a first order of magnitude.

**Calculating GPU memory**
1. Determine the number of parameters in billions; it appears in the name or the model description.
2. Multiply by the memory per parameter: around 2 bytes at 16-bit precision, around 1 byte at 8 bits, around 0.5 bytes at 4 bits (quantisation).
3. Add a margin of about 20 percent or more for the cache of running requests, more for long texts and many concurrent users.
4. Compare the result with the memory of the available graphics cards.

An example shows the range. A model with 8 billion parameters needs around 16 gigabytes for the weights at 16 bits and therefore runs on a single graphics card with 24 gigabytes. A model with 70 billion parameters needs around 140 gigabytes at 16 bits and therefore several data-center graphics cards; at 4 bits the requirement falls to around 35 gigabytes, with a measurable effect on quality that the evaluation tests. On top of the hardware come software for deployment and monitoring as well as a team responsible for updates and security.

## Does the AI Act apply to open-source models?

Yes, the AI Act applies to open models in principle, but it provides relief for models released under free and open-source licenses. This relief no longer applies once a system is used in a high-risk area or for prohibited practices, or touches transparency obligations. For general-purpose models with systemic risk, the full obligations apply regardless of the license. Obligations on copyright and on a summary of the training content also remain for open models.

For companies that build an open model into their own application, one consequence matters most: the obligations arising from using the application follow the purpose and risk of the system and apply regardless of whether the model is open or commercial. The article [EU AI Act](/blog/eu-ai-act/) gives an overview of roles and obligations.

## A portfolio instead of a decision of principle

Companies that sort their use cases by the five criteria usually identify three classes:

- High protection requirements and a clearly defined task: candidate for open models in self-operation.
- Open-ended task and medium protection requirements: candidate for commercial models in a contractually secured environment.
- High volume and a stable profile: candidate for a cost comparison of both options.

A selection rule that fixes these classes prevents every business function from making its own model decisions. Executive management sets guardrails, the business function chooses within them, and the order of applications follows their value contribution. The article [AI management: what executive leadership must steer](/blog/ai-management-what-executives-must-steer/) describes how this steering is built as a whole. The overarching question of whether a company develops, buys or adapts AI is covered in the article [Make or buy](/blog/build-buy-or-adapt-ai/). Model choice is part of that decision and can be revised later, provided the architecture allows for the switch.

## How Niehues Consulting supports model choice

[Niehues Consulting](/) supports model choice as part of AI management and implementation. The approach follows the Niehues Delivery Model with the phases Understand, Build and Implement. In the Understand phase, the requirements, the data classes and the business case are defined for each use case. In the Build phase, a prototype compares both model types in the real process, based on the business function's evaluation cases. In the Implement phase, operating procedures and responsibilities move into regular operation, with internal or external operations teams.

Three of the model's assessment questions matter most for model choice. Does the prototype prove the benefit in the real process with real data? Is the business case confirmed with real values, that is with measured volume and measured cost per request? Who is accountable for operation from day one? Marin Niehues works from the delivery perspective: a model choice counts as made when the model runs in the daily work of the business functions and a named operator is accountable for it. The entry point is the free 30-minute initial consultation via [contact](/#kontakt).

**Recommendation for executive management**
Assess every use case against the same five criteria, review the license before the prototype and evaluate both model types with your own cases. Set a framework that enables switching. The result is a portfolio that balances protection requirements and quality per application and keeps costs controllable.

**Sources**
- Open Source Initiative: The Open Source AI Definition 1.0, 2024
- Regulation (EU) 2024/1689 laying down harmonized rules on artificial intelligence (AI Act), 2024

## Frequently asked questions

### Are open-source models cheaper than commercial models?

With high, steady volume yes, with low or fluctuating volume no: there, commercial models with usage-based billing are cheaper. Open models in self-operation incur fixed costs for infrastructure and staff that only spread out across many requests. The reliable answer comes from the business case for each use case.

### Are open models more secure?

Open models offer more control over the data flow, because operation can be run in your own environment. Security itself results from operation with access protection and monitoring. Commercial providers contractually guarantee a high level of security. The assessment follows the data class of the use case.

### Can a company use both model types in parallel?

Yes, and it is the most common target state. A layer between application and model ensures that applications can switch models without being rebuilt. Executive management determines which class of application may use which model type.

### What does open weights mean for language models?

Open weights means that the provider publishes the trained parameters of the model, so that companies can run and adapt it themselves. This is not automatically open source in the strict sense, because the license may contain restrictions and information about the training data is usually missing. Every selection therefore includes a license review.

### What is the best open-source LLM for business use?

The best open model for business use is the one that solves the business function's cases best in your own evaluation and whose license fits the intended use. The shortlist usually includes models from the Llama, Mistral, Qwen, DeepSeek or Gemma families. For clearly defined tasks, a small model running on a single graphics card is often sufficient.

### Can open-source models be used commercially without license fees?

Models under licenses such as Apache 2.0 or MIT may be used commercially without a license fee, provided license obligations such as attribution are met. Manufacturer licenses such as Llama's also permit commercial use but set their own conditions. In both cases, costs arise for operation, the hardware and the staff.

### Can you run an LLM on your own servers?

Yes, open models can be run on your own servers or in your own cloud environment. This requires graphics cards with sufficient memory, software for deployment and a team accountable for operation, the updates and the security. You can estimate the memory requirement from the number of parameters and the chosen precision.
