About
MAFDET is built by Simone Di Mauro and Simone Di Natale, two friends from Catania who have been programming together since they were kids. Both study at the University of Catania, Computer Science and Computer Engineering. Between them: back end, AI, security, hardware, machine learning. Roughly what it takes to make a language model start working before you press Enter.
Computer Science · Bachelor's, University of Catania
Computer Science, with his ground in back-end development, artificial intelligence and cybersecurity: services and APIs, the models behind them, and what happens when someone tries to break them. He cares about software that runs locally and is safe by default: no accounts, no data leaving the machine, downloads that trust the system's certificates instead of switching checks off.
@MbareSimo99 ↗
Computer Engineering · Bachelor's, University of Catania
Computer Engineering, at home between software design and hardware: how a program is structured, and what the machine actually does with it. His interests run through machine learning, automation and performance: fitting a model into a 6 GB card, measuring what every layer costs, and automating the benchmarks so a result can be repeated instead of believed.
@TheLastOxygen ↗◆ In practice they do everything together: the engine, the app, the research, the arguments about all three. The areas above are where each one starts from, not a division of labour.
The story
They grew up in Catania and started programming together as kids: first games and scripts, then anything that needed a computer to do something it did not do out of the box. The habit stuck. University gave it two names, Computer Science for one and Computer Engineering for the other, both at the University of Catania.
The two curricula overlap exactly where MAFDET lives. On one side the theory of how a language model reads (prefill, parallel and fast) and writes (decode, one token at a time). On the other the hardware that has to do it fast enough to feel instant, with a KV cache that fits in six gigabytes.
MAFDET started from a measurement, not a slogan. A local model on an RTX 2060 answers a word problem about twelve seconds after Enter, and most of those seconds are reasoning. During the twenty or thirty seconds spent typing the question, the GPU is idle. The question followed by itself: can that time be used?
Reading ahead was the easy half: keep the KV cache in sync with the text, hold back the last tokens, trim to the common prefix on every edit. It is exact, and mafdet judge proves it against a from-scratch computation. Thinking ahead was the hard half: a model reasoning on a message that is not finished, with its notes kept across edits. That is not exact, so it was measured the only honest way, on problems with a known answer, dozens of times, on two machines.
It was built on purpose for the hardware most people own: a six-gigabyte card from 2019 as the reference machine, an M2 Pro as the second, a CPU path for everything else. Faster hardware makes it faster. It does not need it.
How they work
Measure it
Every claim on this site has a script and a JSON file behind it in the repository. When something did not work, the gatherer, the speculative drafts, it is written down with the numbers, not deleted.
Say what it is
A prototype is called a prototype. Linux is untested and the site says so. Reading ahead is exact and thinking ahead is not, and the difference is explained rather than smoothed over.
Build for real machines
No PyTorch, no cloud, no account. The app sets itself up on the GPU you have, or on the CPU if that is all there is, and never sends a message anywhere.
The best place is GitHub: an issue, a benchmark from your machine, a pull request. If you have an AMD Radeon, they would especially like to hear from you.