AI models developed in Cyprus (CM-01.0)

CM-01.0 · COMPANIES

AI models developed in Cyprus

estimateCurated, reviewed 7 Aug 2026

AI models of one billion parameters or more, trained from scratch or as a full fine-tune, whose publisher releases the complete weights under its own name. The publisher must have its registered office in Cyprus, or the training must have been carried out on compute located in Cyprus. Adapters, delta weights, API-only releases and third-party mirrors do not count. A fine-tune of another publisher’s base model counts only if the base model’s licence allows the full merged weights to be published.

Register reading

A single reading. The source publishes its current state only.

Period: 2026 register
Read on: 7 Aug 2026
From: Hugging Face model index

The register does not keep a history. Rebuilding past values from announcements and press coverage would not be reliable, so no series is published. The figure is updated when the register changes.

estimateTier C estimate

The figure of zero is a negative claim based on open registers. No register is complete, and a model can exist without being listed, so the figure is tier C rather than tier A. A model counts only if it passes all three of the following tests. Scale: it has one billion parameters or more, as disclosed in the model card, the accompanying paper or the released weight files, and a fine-tune is counted at its full merged size. Attribution: the publisher of record has its registered office in Cyprus, which can be checked against the Registrar of Companies, or the training run is documented as having taken place on compute physically located in Cyprus. Weights: the publisher releases a complete, standalone set of weights under its own name, whether the licence is open or access is gated. Adapters, delta weights, API-only releases and third-party mirrors fail this test, and a fine-tune of a third-party base model passes only if that model’s licence allows the publisher to redistribute the full merged weights. On 7 August 2026 we searched the Hugging Face model index for the names and handles of Cypriot universities, research centres and public research bodies, and separately for models tagged to Cyprus. The search found fine-tuned adapters and small task-specific checkpoints published by individuals, but no model of qualifying scale published by a Cypriot organisation. On 27 August 2026 we widened the attribution and scope criteria: full fine-tunes now count alongside pretraining, and a qualifying training run on compute located in Cyprus counts wherever the publisher is registered. Every result recorded from the 7 August search still fails at least one test under the wider criteria. The adapters fail the weights test, the small checkpoints fail the scale test and the individual publishers fail the attribution test, so the figure remains zero. A new search under the wider criteria is due at the next review, and until then the figure should not be cited as current under the new definition. We also checked publicly announced national and EuroHPC compute allocations for a qualifying training run and found none. We will revise the figure if we are notified of a qualifying model, and any such claim can be tested against the criteria set out here.

This figure should not be cited as an official statistic. Read the method.

Provenance

Source: Hugging Face model index
Publisher: Hugging Face
Series: CM-01.0 · irregular
Period: 2026 register
Quality: Tier C — estimate
Value origin: Last reviewed figure
Next release: Not announced
Licence: Hugging Face terms of service

Download CSV API endpoint

REFERENCE BAND

0 → 1score 0/100

We set the ceiling at one. The question is whether AI developers in Cyprus produce and publish models at this scale at all, whether trained from scratch or built on another publisher’s base model, rather than only using models made elsewhere. A second such model would tell us much less than the first.

NO PEER COMPARISON

This figure is compiled from primary documents, and so far we have read them for this country only. A comparison with part of the peer group would give a misleading rank, so none is shown. How the peer group is chosen.

AI models developed in Cyprus

Register reading

Provenance

REFERENCE BAND

NO PEER COMPARISON

RELATED METRICS