RFGPT: the AI model that reads radio signals in plain language opens to telecom operators at MWC Doha

Radio spectrum is invisible, crowded and, until now, readable only by specialists with expensive equipment and years of training. RFGPT is an attempt to change that: a language model that takes in raw radio signals and answers questions about them the way a chatbot answers questions about text. From Sunday, the people who actually run mobile networks get to test whether that promise holds up in practice.

What RFGPT actually does#

RFGPT works by turning radio signals into pictures a language model can read. A receiver captures an IQ waveform, the system converts it into a spectrogram — an image showing how signal energy changes across time and frequency — and encodes that image into RF tokens that are injected into a decoder-only large language model. The result is a model that can be asked, in plain language, what is transmitting, on which technology, and whether signals are interfering with each other.

The model identifies six wireless technologies — Wi-Fi, Bluetooth, LTE, 5G New Radio, UMTS and DVB-S2 — and distinguishes uplink from downlink transmissions where that applies. It recognises modulation schemes, analyses wideband environments with several simultaneous signals, and finds time-frequency overlap between transmissions to estimate interference. It can also estimate wireless activity by counting WLAN and NR users in an observation window, and it interprets 5G NR parameters including subcarrier spacing, synchronisation signal block patterns, and CSI-RS and SRS configurations.

“We gave a language model its first glimpse of the electromagnetic spectrum and the view is already remarkable,” said Professor Merouane Debbah, Senior Director of Khalifa University’s Digital Future Institute, when the model was first announced in February. The project team includes postdoctoral fellows Hang Zou and Yu Tian, research scientists Lina Bariah and Samson Lasaulce, Chongwen Huang and PhD student Bohao Wang.

What the trial gives operators#

Participating operators get two weeks of API access to RFGPT, running on AMD compute. To join, an operator submits a use case to the research team, which then agrees the evaluation scope, access arrangements and trial dates with each operator. The stated test areas are network planning, spectrum management and interference analysis, plus radio access network optimisation and network operations and performance management.

Operators will be asked to feed back on performance, usability and relevance to their own applications. The GSMA says the results will show where the model adds value and what it needs to go further. The trial also doubles as a showcase for AMD’s AI compute infrastructure: the chipmaker’s data-centre revenue rose 97% in its third quarter, part of a record $13.4 billion quarter, up 45% year on year. AMD’s pivot toward the data centre is the same shift its chief executive described last week as the company becoming “data center first.”

The trial is the latest step in a partnership between the GSMA’s innovation arm, GSMA Foundry, and Khalifa University that began in November 2025, aimed at producing open telecom AI assets and benchmarking tools. Telecom AI is a crowded lane: Calix released an 8B telecom model on Hugging Face in October, and China Telecom and HKUST opened a joint AI-for-telecoms lab the same month. RFGPT is different in kind — it reads the airwaves themselves, not network documentation.

The numbers behind it#

RFGPT was trained on about 625,000 computer-generated radio signal examples drawn from roughly 12,000 simulated scenes covering the six wireless technologies, with no manual labelling. In the university’s benchmark tests it outperformed existing baseline models by up to 75.4% on radio-frequency spectrogram tasks, and correctly counted the number of signals in a spectrogram about 98% of the time — a task general-purpose AI models almost never manage, according to the university. It also beat OpenAI’s GPT-4 and Alibaba’s Qwen 2.5 on modulation classification and 5G NR information extraction.

Those figures are vendor-reported: Khalifa University ran the evaluations itself, and no independent reproduction has been published. That caveat matters more than usual here, because the gap between a controlled synthetic benchmark and a live network is exactly what the operator trial is designed to measure. The same caution applies to benchmark claims across the industry — SemiAnalysis’ GDPx benchmark, launched today, found the best AI agent finishes just 12% of real professional work, a reminder that lab numbers and real-world performance are different things.

Why it matters#

Today, understanding what is happening in the spectrum means sending a radio-frequency engineer with a spectrum analyser. If a language model can do the first pass — spotting an interfering transmitter, identifying an unknown signal, counting users in a cell — network planning and troubleshooting get faster and cheaper, and spectrum authorities get a tool that speaks their language instead of requiring them to learn the model’s.

The longer game is 6G. Khalifa University frames RFGPT as a step toward AI-native radio systems, where RF perception directly supports network optimisation and policy decisions. The GSMA’s MWC Doha showcase lists RFGPT as an “AI-powered spectrum-intelligence platform” alongside PhoneGPT, an agent that understands a user and their smartphone context — a sign that the industry body sees spectrum intelligence as a showcase technology, not a lab curiosity.

What to watch#

Three things. First, what operators actually report: whether two weeks of API access produces concrete findings on real networks, or whether the model’s synthetic training shows its limits outside the lab. Second, whether the trial leads to wider access — the university has not said what happens after the trial window, or whether the model will be opened more broadly. Third, whether the GSMA publishes the results: the association says the trial will show where RFGPT adds value, and a public readout would be the first independent signal on whether a language model can really read the airwaves.

Until then, the fair description is narrow: a university model with strong self-reported numbers, now in the hands of the people whose networks it has to understand.

Sources: Middle East AI News (November 8, 2026; source: Khalifa University); Khalifa University’s RF-GPT announcement; the GSMA’s MWC Doha 2026 page. Benchmark figures are the university’s own, not independently verified. The images are AI-generated illustrations.