Does Every UAE Company Need Its Own AI Model?

Analysis | UAE researchers are releasing models that businesses can evaluate and adapt. The case for owning model weights starts with a specific capability or operating constraint that a simpler system cannot meet.

A Dubai services company wants an assistant that can understand Emirati Arabic, answer from the latest service manual and work within a controlled network. These requirements may sound like a single argument for building a model. In practice they point to three different decisions. The language requirement concerns how well the system understands users. The manual concerns where the assistant gets its facts. The network concerns where the service runs and who has access to it. A company can make each decision separately and test whether they need to change the model at all.

This matters because the phrase our own model can refer to anything from a dedicated deployment of existing weights to a foundation model trained from scratch. The operational responsibilities differ sharply. Before funding model development, name the failure that remains after testing the available models in the intended workflow. Then identify which part of the system would have to change to fix it.

A UAE foundation model can be a starting point

In September 2026, the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence released K2 Horizon, a family of six models ranging from 0.9 billion to 375 billion parameters. Its technical release describes publishing weights, configurations and intermediate checkpoints alongside training data or construction recipes where redistribution was unavailable. That last distinction matters to a team assessing reproducibility. Instructions for assembling a dataset do not automatically give a business the underlying material or rights to use it. IFM’s K2 Horizon technical release

MBZUAI’s K2 Horizon Instagram reel introduces the family and explains the institute’s approach to open research. The technical release is where a prospective adopter can look for the differences among models, the published artifacts and the terms that apply to them.

A month later, MBZUAI and Inception42 announced Jais 3, a set of Arabic-focused models built on K2 Horizon. That provides a UAE example of building a new capability on a released foundation. It does not make a research programme of that scale the right template for an ordinary business. The relevant lesson is that the starting point and the finished capability can be different. A UAE company may be able to evaluate a foundation already adapted for its language requirements before commissioning a new adaptation. MBZUAI’s Jais 3 announcement

Abu Dhabi’s Technology Innovation Institute offers a separate example. In October 2026 it introduced Falcon-Emirati, a seven-billion-parameter language model designed for Emirati Arabic, alongside models for speech recognition and reading text from images. A contact centre that works with recorded speech has a different test from a team that receives written messages. It may need to test transcription and language understanding separately. A model that handles a written phrase correctly after a clean transcript can still be part of a system that mishears the phrase in a call. TII’s Falcon-Emirati, speech and visual text announcement

Find out what failed before changing the weights

Imagine the services company asks its assistant whether a maintenance appointment can be moved. One answer uses an obsolete policy. Another uses the current policy but misreads a colloquial request. A third identifies the correct rule and offers a date that the booking system cannot deliver. A single accuracy score would hide the fact that all three errors need different remedies.

For the stale policy, inspect what the system retrieved. If it supplied last year’s manual, fix document ownership, effective dates and search filters first. Retrieval-augmented generation gives the model relevant material at the time it answers. It lets the business revise the document separately from the model, provided the search index and document permissions keep up. Fine-tuning changes learned behaviour through training on examples. It can be useful for a stable task, but a company that uses it to encode regularly changing policy still needs a way to replace stale information and verify the update. Microsoft’s technical guidance on retrieval and fine-tuning

For the misunderstood request, give the model the correct policy passage and see if the error persists. Test the same meaning expressed in Arabic, English and the mixed wording that users actually send. If the problem persists after correcting the input and simplifying the task, compare existing models on those cases. Fine-tuning becomes a plausible option when the errors are consistent, the business has representative examples it can use, and a held-out set can show whether the adaptation improved those errors without introducing worse ones elsewhere.

For the unavailable appointment, validate dates in the booking system and constrain the action to available slots. The company can ask the model to interpret a message while keeping the booking transaction under ordinary application controls. Training a model to produce more plausible dates would leave the underlying problem intact. The workflow still needs a current availability check and a clear rule for what happens if that check fails.

Hosting and adaptation are separate choices

A company can buy access to a managed model or operate a deployment it controls, subject to the terms of the model and infrastructure it chooses. It can also run an existing model with its original weights or use an adapted version. A customer’s requirement that prompts stay within an approved environment may affect hosting and access. It does not by itself establish that the company has to create model weights.

Draw the data path from the employee’s device to the answer and onward to any records. Include the search index, logs, backups, monitoring, support access and tools the model can call. A private model endpoint can still be part of a system that sends transcripts to a separate service. A deployment description should cover those connections before the company treats the boundary as settled. Check the exact model release and the terms for the planned use, especially if a business intends to provide a model-based service to other organisations.

Operating the deployment also means taking responsibility for capacity, recovery and updates. Measure response times during the busiest part of the day. Check what happens when the service loses a machine and who can restore the previous model if a change causes errors. A dedicated deployment can offer useful control when the team can maintain it. A smaller model can be easier to serve, but that advantage only matters if it can still do the job at the required quality.

Pay for an outcome the company can verify

Compare alternatives on the same work sample, with the same documents and the same opportunities for human review. Include ordinary cases, difficult language, conflicting policies and requests that should be refused or escalated. Report the hardest categories separately. A system that writes excellent routine answers but repeatedly mishandles one important class of exceptions may need to stay out of that part of the workflow.

For a hypothetical cost comparison, suppose a UAE team handles 20,000 cases a month. With a managed model, 12,000 pass the quality gate without rework and monthly model, infrastructure and review costs total AED 24,000. A self-hosted alternative produces 15,000 such cases and costs AED 36,000 on the same basis. That is AED 2.00 versus AED 2.40 per case that passes without rework. A higher number of acceptable cases can justify the second option if the extra 3,000 cases have enough operational value or if hosting solves a requirement the managed option cannot meet. The numbers are illustrative. A full comparison should also show the cost of handling failures and any transition costs excluded from the monthly figures.

Keep a record of what the team would have to redo when the model changes. A maintained test set, approved source documents and documented interfaces can carry over to a new deployment. A tuning dataset may also remain useful if its permissions and labels are clear, though a new base model will need fresh evaluation. This matters even when the current system works. A team that can replace one component without retraining reviewers or rebuilding its records has more room to respond when costs, support or model behaviour change.

Pretraining a foundation from scratch calls for a stronger case. A company would need a capability it cannot reach through existing models and adaptation, enough usable data and compute, and people who can maintain the result over time. If the proprietary advantage comes from a service manual, customer workflow or operating dataset, much of the value may sit in the surrounding system. The company can retain that advantage while replacing the model as better candidates become available.

A useful decision record says which constraint the current option solves, which errors remain and what result would justify changing course. Start with the simplest candidate that meets the deployment boundary, measure it on UAE work, and change one layer at a time. Train or adapt a model when the company can show that model behaviour is the bottleneck and has a reliable way to tell whether the investment removes it.