In the midst of this transformation, a critical question arises for any company: what type of AI model should it adopt?
The key difference, in brief. A closed (or proprietary) AI model is one developed and controlled entirely by a provider, accessed through an API or SaaS platform. An open-weight model is one whose parameters can be downloaded, run and fine-tuned in your own environments. The choice between them depends on three axes: cost, control over data and in-house technical capability.
An important note on terminology. The industry frequently uses the term "open source" to refer to models such as Llama (Meta), Mistral, Qwen (Alibaba) or DeepSeek. However, the Open Source Initiative (OSI) formally rejects that usage: the Llama licence imposes restrictions that do not meet the official definition of open source, including a limit on free commercial use for companies exceeding 700 million monthly active users. The technically more accurate term is "open-weight models". This distinction matters when assessing legal and compliance obligations, and it also has a little-known implication: the multimodal versions of Llama 4 expressly exclude individuals and companies domiciled in the European Union from their licence, a restriction Meta links to the regulatory uncertainty of the EU AI Act. For a Spanish company considering Llama as an option, this is a point worth checking before moving forward, depending on the specific model and version.
Have open models closed the performance gap? In 2023, proprietary models clearly outperformed open-weight ones in the main benchmarks, with a lead of more than a year in some tests. That advantage has narrowed considerably: according to the most recent analysis by Epoch AI, by mid-2026 the most advanced open models lag behind the closed state of the art by an average of around four months, equivalent to some eight points on its capability index (ECI). In structured reasoning and coding tasks, some open models such as GLM or DeepSeek-V4 already match or surpass proprietary models in specific tests.
This narrative should be qualified with a less frequently cited fact: not all of the savings from open models materialise in practice. An analysis by the firm Nous Research of nineteen models found that some open models consume between 1.5 and 10 times more tokens than closed ones to solve the same task, which can offset, wholly or partly, their lower cost per token. The real saving depends on the use case and must be measured, not taken for granted.
Closed models: power under lock and key
Closed models are developed and managed entirely by a provider, which controls the whole cycle: training, infrastructure, updates and access. Representative examples are ChatGPT (OpenAI), Gemini (Google) or Claude (Anthropic).
Advantages
- Performance optimised for general use. The major providers train on massive cloud infrastructures, producing robust models that are ready to use without additional configuration.
- Ease of access. You only need to connect to an API or SaaS platform to start using them, with no need for a specialised technical team.
- Continuous improvement managed by the provider. Updates reach the user automatically.
Limitations
- Total dependence on the provider. The company is subject to pricing, changes to terms of use and service availability.
- Variable costs that scale with use. Pay-per-token pricing can grow significantly as the number of queries increases.
- Less control over data. Information is processed on external servers, which can raise issues in regulated sectors.
- Limited customisation. Adapting the model to the particularities of each business is more complex and in many cases not available.
Open-weight models: flexibility and data sovereignty
Open-weight models make it possible to download the parameters, run them in your own environments and fine-tune them with internal data. Notable examples are Llama (Meta), Mistral, Qwen (Alibaba), DeepSeek and gpt-oss, the open-weight model that OpenAI itself released as an alternative to its closed offering. As noted above, not all of these models are technically "open source" according to the OSI, and some, such as the multimodal versions of Llama 4, have specific geographical restrictions that should be reviewed case by case. Even so, all of them allow a far greater degree of control than proprietary models.
Advantages
- Full control over where data is stored and processed. This is critical in finance, healthcare and public administration, where regulations require that information does not leave the corporate perimeter.
- Deep customisation. The model can be fine-tuned with internal data and adapted to specific business processes using fine-tuning techniques.
- Potentially lower cost at scale, though not automatically. An analysis by IBM and Morning Consult found that 51% of companies using open models reported a positive return on investment, compared with 41% of those using proprietary models. As explained earlier, that saving depends on managing token consumption and infrastructure well; it is not automatic simply because an open model is used.
- A fast-moving innovation ecosystem. The community, as well as labs such as Moonshot AI, Alibaba or Zhipu AI, contribute improvements, extensions and specialised models at great speed.
Limitations
- Greater technical complexity. Deploying and maintaining an open-weight model requires teams specialised in MLOps.
- Initial investment in infrastructure. Computing capacity is needed, whether on-premise or in a private cloud.
- Responsibility for security lies with the company. There is no single provider guaranteeing operation or offering an SLA.
The three decision axes: cost, security and control
Before choosing, it is worth assessing the company's situation on these three axes.
Cost. Closed models have a low initial investment and are ideal for getting started quickly, but their variable costs grow with use. Open models require more initial investment in staff and infrastructure, but are more predictable and scalable in the medium and long term if consumption is managed well. According to a study by Frank Nagle and Daniel Yue for the Linux Foundation, the unrealised value resulting from the under-use of open models in companies amounts to 24.8 billion dollars a year.
Security. With closed models, security depends on the provider's policies and data travels to external environments. With open models, the company keeps the data within its perimeter, but takes on responsibility for implementing its own security controls, audits and regulatory compliance.
Control. Closed models are easy to integrate, but offer little scope for customisation. Open models allow the model's behaviour to be adapted to business processes, albeit with greater management effort.
Strategies by type of company
- Start-ups and SMEs. They look for speed and low initial cost. Closed models are usually the best option for validating ideas and launching projects without large investments in infrastructure.
- Medium-sized and large companies. They handle significant volumes of sensitive data and need to differentiate themselves with tailor-made solutions. Open models provide greater control and scope for customisation.
- Regulated sectors (finance, healthcare, public administration). Data sovereignty is a priority. Open models are preferable, always combined with audits and regulatory compliance. Red Hat documents this trend in its analysis of the state of open models, noting that sectors with strict data residency requirements are among the first to adopt locally deployed models.
- Environments with a wide variety of use cases. The ideal path is a hybrid architecture that combines the speed of closed models with the flexibility of open ones.
The hybrid approach: the best of both models
More and more organisations are opting for hybrid architectures, in which closed and open models coexist. Closed models are used for general, creative or quick-support tasks, while open-weight models are applied to critical processes where data control and customisation are decisive.
This approach makes it possible to optimise costs by using the most suitable model in each case, keep the most sensitive data secure and scale flexibly as business needs evolve.
Practical examples
- Customer service. A closed model can handle frequent queries, while an open model connected to RAG architectures (LurNova's RAG solutions) answers questions based on internal documentation and proprietary data with precision.
- Finance. An open model can analyse sensitive internal data under strict control, while a closed one can generate summary reports for general use.
- Industry. A hybrid environment allows closed models to be used for general process optimisation and open models for predictive analysis with critical production data that cannot leave the corporate environment.
At LurNova we help you design the right ecosystem
At LurNova we know there is no single valid answer. That is why we have developed the LurNova AI Model Hub, an architecture that integrates private models and open models in a single environment, with intelligent routing that decides which model to use in each case, cost control by optimising consumption, advanced security that keeps sensitive data under control, and full flexibility to combine rapid innovation with deep customisation.
The future of corporate AI is not about choosing between open or closed, but about designing an ecosystem that makes the most of both approaches.
CEO of LurNova, with extensive experience in technology and digital transformation. Involved in developing technology projects and applying artificial intelligence in organisations, he combines technical knowledge with business vision.
