System requirements
Alyx runs the language model on your computer. Nothing about your project is sent anywhere, and that is the whole point — but it means the model has to fit in your machine, next to Revit. This page tells you exactly what fits where, so you can check before you buy.
The short answer
It depends on two things, not one: your graphics card, and how heavy your Revit project is. Revit uses the same card, and a house uses a few hundred megabytes of it while a hospital uses several gigabytes.
| Your graphics card | With a light project | With a heavy project |
|---|---|---|
| 16 GB or more | Alyx Large, wide context | Alyx Large — the recommended setup, and the one Alyx is developed on |
| 12 GB | Alyx Large | Alyx Medium |
| 8 GB | Alyx Medium | Alyx Small |
| 6 GB | Alyx Small | Processor only |
| 4 GB or less, integrated graphics, or no NVIDIA card | Processor only, and only with 32 GB of RAM | Same |
Three things this table is careful about:
- Revit itself runs happily on far weaker graphics than these. It falls back to its own processor pipeline and keeps working. These figures are about the AI model, which is a different and much larger tenant of the same memory. If your machine runs Revit today, it will keep running Revit.
- Alyx measures your machine on first run and recommends a model, telling you which numbers it decided on. You can pick a different one right there, and change it later.
- That recommendation is deliberately more cautious than this table. It is made once, before you open any project, so it assumes the heavy case and reserves ~3 GB for Revit. That is why an 8 GB card is recommended the processor, even though the table says a GPU tier would fit with a light project — you can choose it yourself, and if your projects are light you should. Recommending too big is the one error a click cannot undo: on Windows an over-large model does not fail, it drags into system memory and sits under a token per second, taking Revit’s performance with it.
Why Revit is counted in
Because Alyx does not get the graphics card to itself. It lives inside Revit, which is already holding your building model in that same memory. A page quoting your card’s total memory would be technically true and practically misleading — and you would find out after paying.
The full table
| Minimum | Recommended | |
|---|---|---|
| Operating system | Windows 10 21H1 (10.0.19041) 64-bit | Windows 11 |
| Host application | Autodesk Revit 2027 | Autodesk Revit 2027 |
| Graphics card | NVIDIA with 6 GB dedicated VRAM (Alyx Small) | NVIDIA with 16 GB or more (Alyx Large) |
| System memory | 16 GB | 32 GB |
| Free disk space | ~10 GB | 20 GB |
| Internet | To download the model and activate the licence, then periodically to renew permission to use it | Working on your model is always offline |
What each model costs you
| Download | Graphics memory while running | Image understanding | |
|---|---|---|---|
| Alyx Large | 6.65 GiB | ~9.7 GB default; ~11.0 GB at a 32,768-token context | Yes |
| Alyx Medium | 4.90 GiB | ~7.5 GB with images; ~6.4 GB without | Yes |
| Alyx Small | 2.49 GiB | ~4.7 GB | Not included |
You only download the model you choose. You can change your mind later and delete the one you no longer want.
Why Alyx Small does not understand images
Because the component that does does not shrink with the model: it is 940 MB on the small tier and 944 MB on the medium one. Alyx Small’s entire saving is in the language model (1.5 GB less), and loading that component would hand most of it back — with it, the tier would not fit the machines it exists for. So it is neither downloaded nor loaded. If you need Alyx to read screenshots and drawings, Alyx Medium is your tier.
Does DDR5 memory help?
It depends entirely on whether the model fits in your graphics card, and the answer surprises people.
- If the model fits in the card — the recommended setup — DDR4 and DDR5 make almost no difference. The model is already in the card’s own memory and system memory is barely involved. Alyx is developed on a DDR4 machine running the large model at a 32,768-token context.
- If you have no compatible card, memory speed becomes the bottleneck and DDR5 is roughly 30-50% faster than DDR4. That is a real gain, but it applies to a path where the first reply already takes minutes.
DDR5 does not reduce how much memory you need. Capacity and speed are different things: the figures above are about fitting the model and Revit at once, and a faster module of the same size fits exactly the same amount. If it says 32 GB, it says 32 GB in DDR4 and in DDR5.
And one thing matters more than the generation — the one people miss when buying a laptop: use two modules, not one. Two sticks run in dual channel and beat a single stick of the same total capacity by a wide margin. A laptop with a single DDR5 module can be slower than one with two DDR4 modules. Plenty of mid-range laptops ship with one module and an empty slot.
If you have no compatible graphics card
Alyx still runs, on the processor, and it needs 32 GB of system memory to do it. Be aware of what that means in practice: replies come at a usable pace once a conversation is going, but Alyx reads a fixed instruction set before its first answer of each session, and on a processor that costs roughly one to three minutes. We would rather you know this now than discover it after buying.
How to check your own machine
- Press
Ctrl + Shift + Escto open Task Manager, go to Performance and select your GPU. - Read “Dedicated GPU memory”. That is the number this page refers to. Shared memory does not count — it is system RAM, and it is far too slow for this.
- For memory: under Performance → Memory, look at “Slots used”. If it reads 1 of 2 you are running single channel, and adding a second matching module is the cheapest speed you can buy.
- Or simply install Alyx and start it: the trial measures your machine and tells you what fits before downloading anything.
The download figures are derived from the real file sizes, not from estimates, and a product test pins them against the signed manifest. The graphics-memory figures are calculated from those sizes and corroborated against a real measurement. The reference machine is an NVIDIA RTX 5060 Ti with 16 GB and DDR4 memory, which runs Alyx Large with Revit open at a 32,768-token context.