Open Models: A Way Forward in the AI Era

When we think about the future of AI, we think about the people who will have to live and work with it. The researcher with a small budget. The independent business owner handling sensitive documents. The student who needs a useful explanation, not another subscription they can barely justify.

For them, the important questions are fairly ordinary. Does this help me? Can I afford to keep using it? Who handles my information? What happens if the company behind it changes its mind?

Those questions are part of why we are building Underleau around open models. Price matters to us. So do independence, accountability, and the possibility of a future in which useful intelligence is something many people can build on, rather than something a handful of companies alone get to distribute.

Underleau is still in development. This is the reasoning behind our model strategy, not an announcement that production AI is already available.

First, what do we mean by open?

A model's weights are the learned numerical parameters that allow it to produce an answer. When those weights are released, other people can download and run the model under its licence, rather than only access it through its creator's service.

That is a meaningful freedom. It is not the same as having the complete recipe for making the model.

Training data, training code, documentation, and permissions vary substantially between releases. The Open Source Initiative's definition of open-source AI goes beyond access to finished weights: it also requires code, data information, and freedoms to study, change, use, and share the system. Some projects publish much more of that material than others.

For accuracy, we describe our intended foundation as open-weight models, while supporting the broader open-source movement. We will review individual releases and their licences. A download button alone does not make a model unrestricted, suitable for commercial use, or appropriate for Underleau.

Affordability is part of usefulness

A tool can be impressive in a demonstration and still be too expensive to use comfortably in ordinary work.

Understanding a difficult document often takes several attempts. You compare interpretations, notice something missing, ask a follow-up, and revise your conclusion. If every step is costly, the product starts discouraging the very exploration it is supposed to enable.

Open weights give different providers the opportunity to serve the same model, subject to its terms. That creates room for competition over price, speed, and how the model is operated. Smaller models also give us more ways to match the amount of computing to the actual task. We see that flexibility as a practical advantage for building an affordable product.

It does not mean open models are always cheaper. Closed providers can offer excellent prices, and running a model ourselves would still mean paying for hardware, operations, and expertise. A low token price is also a poor bargain if the answer is unreliable and needs three more attempts.

Our aim is to reduce the cost of a useful result, not merely the cost of generating words. That means selecting capable models, giving them relevant evidence, reusing work where appropriate, and testing the whole workflow. On a canvas full of documents, sending only the material relevant to a question matters as much as choosing an economical model.

Affordability should come from good engineering. It should never require quietly accepting worse answers or weaker safeguards.

Sovereignty means having meaningful choices

Sovereignty can sound abstract until you depend on a service you cannot replace.

With a closed model available only through its creator's API, access depends on that provider continuing to offer it on terms that work for you. Prices, availability, and model behaviour can change. A product built around one irreplaceable service inherits those dependencies.

Open weights offer another route. Where the licence permits, a model can be served by different operators or run on infrastructure chosen by the organisation using it. That gives builders more room to change providers, assess regional hosting options, and preserve access to a particular release.

For example, Mistral's January 2025 Small 3 release made pretrained and instruction-tuned checkpoints available under Apache 2.0, with both local deployment and hosted access routes. It is one concrete example of a model being available beyond a single API.

That freedom has limits. Changing providers still takes integration work and testing; the same weights do not guarantee identical outputs, speed, or service quality. Licences can impose restrictions. Hardware and cloud infrastructure remain dependencies.

Nor do public weights automatically keep private data private. The operator running a model still matters: where processing happens, what is retained, and what its contracts permit. A model's country of origin and the location where a request is processed are different questions.

Our production plan is to use reviewed hosted endpoints, not to promise offline AI or self-host every model at launch. EU processing and our commitment never to use customer data for model training must be carried through the actual services we select. Openness gives us more choices; it does not do that work for us.

Shared work gives more people a starting point

The people moving open AI forward are not only the laboratories releasing checkpoints. They are also researchers evaluating failures, engineers making models run more efficiently, translators improving language coverage, and independent developers finding uses the original creators did not anticipate.

When work is available to build on, a useful improvement can travel further than one product.

Projects such as Ai2's Olmo show what a deeper commitment to openness can look like, sharing model weights alongside training data, code, and evaluation tools. That gives others more material with which to investigate how a system was built and test ideas of their own.

We value that possibility. A wider range of people should be able to ask how these systems work and participate in improving them. Access to a checkpoint is not complete explainability, but access to more of the underlying work creates opportunities that a closed interface cannot provide on its own.

As a small team building on that ecosystem, we should be candid about the debt we owe it. Giving credit and linking original work are basic responsibilities. We also should not claim contributions, partnerships, or research achievements we have not made.

Quality still has to be earned

None of this requires pretending that closed models are bad. Many are exceptionally capable. Nor does it mean that every open model can match every closed model on every task.

For Underleau, the test is whether a model can do the work reliably: understand the relevant sources, distinguish evidence from inference, use tools correctly, and acknowledge when the answer is uncertain. A benchmark can help us decide what to investigate. It cannot establish that a model will handle someone's actual documents well.

Our planned approach is to evaluate models against representative workflows, including difficult and adversarial cases. The surrounding product matters too: retrieval, permissions, source references, constrained tools, and safety checks all contribute to the result. An open model does not get a lower standard because we like the principle behind it.

Openness is not a safety certificate. Open and closed models can both make mistakes or follow malicious instructions hidden in material they read. The application must enforce boundaries independently of a model's answer, especially when private documents and connected apps are involved.

A direction we are proud to take

Choosing open models is a practical decision and a statement about the kind of industry we want to help build. We want capable AI to be affordable, adaptable, and less dependent on permission from a single supplier. We want the people sharing the underlying work to receive credit for making that possible.

That support is not a blanket endorsement of the companies behind these models, a partnership with them, or approval of their conduct. It does not excuse actions that conflict with Underleau's values. We can recognise the value of a release while remaining critical of its creator's decisions, and we remain responsible for what we choose to deploy.

Our Open Models page introduces some of this ecosystem. It is a showcase, not a list of models already running in Underleau. Production selection, provider review, and testing are still ahead.

The ambition is straightforward: build a useful thinking tool on foundations that leave more room for others to build, too.