Khalifa University Opens Its RFGPT Radio Signal Model to Telecom Operator Trials

Operators that submit a use case can get two weeks of API access to the model, which runs on AMD computing hardware. The research team then agrees the scope of the evaluation, access arrangements and trial dates with each operator. Participants will be asked for feedback on performance, usability and how relevant the model is to their own networks, which the GSMA says will show where the model adds value and what it needs to develop further.

The trial moves a piece of Gulf-developed telecom AI from the lab into the hands of the companies that run networks. Interpreting radio frequency data has traditionally needed specialist engineers. RFGPT is designed to let staff ask questions about signals in plain language, which its developers say could make network planning and troubleshooting faster and cheaper.

RFGPT was built by the university's Digital Future Institute and first announced in February. The researchers describe it as the first radio frequency language model to feed RF spectrograms, visual maps of signal energy over time and frequency, into a multimodal model, so that one system can reason over wireless signals and text. Earlier telecom AI typically relied on a separate deep-learning model for each signal-processing task.

The model identifies Wi-Fi, Bluetooth, LTE, 5G New Radio, UMTS and DVB-S2 transmissions and, where relevant, tells uplink from downlink. It recognises modulation schemes and can analyse wideband scenes with several signals at once, the kind of crowded spectrum where operators need to know what is transmitting and how. For coexistence analysis it finds overlaps between transmissions in time and frequency and estimates the strength of interference. It can also estimate Wi-Fi and 5G user densities in an observation window and interpret 5G parameters such as subcarrier spacing and synchronisation signal block patterns.

The team trained RFGPT on about 625,000 synthetic instruction examples drawn from roughly 12,000 simulated scenes covering six wireless technologies, with no manual labelling. In the researchers' benchmark tests, it outperformed OpenAI's GPT-4 and Alibaba's Qwen 2.5 on tasks including modulation classification and extracting 5G information.

The operator trial builds on a strategic partnership between GSMA Foundry, the association's innovation arm, and Khalifa University that began in November 2025 to produce open telecom AI assets and benchmarking tools. RFGPT is also one of the models in the GSMA's Open Telco AI initiative, launched at MWC Barcelona in March with AT&T and AMD as founding supporters.

For AMD, the trial is a chance to put its AI accelerators in front of operators. The company reported last week that third-quarter revenue reached a record $13.4 billion, up 45% from a year earlier, with data centre revenue up 97% on record sales of its Instinct AI GPUs and EPYC processors.

Operators can apply through a form published by the research team; the university has not said how many have signed up or when results will be released.