There is nothing to set up. This page is for the first download and for everything you might want to change afterwards.
Getting it running
Unzip the download anywhere and run the blucta.exe. There is no installer, nothing is written to the registry, and nothing else has to be installed first — the parts that run the models ship inside the program.
The first run
On the first start, Blucta AI checks what your PC can handle, chooses its models, and offers to download them. About 19 GB in total:
| What | Its job | Size |
|---|---|---|
| Qwen3.8-27B | Reads the screen, decides every step, and points at what to click | 17.4 GB |
| bge-m3 | Finds the right skill and the right old note | 0.6 GB |
| Whisper large-v3-turbo | Voice typing, in any language | 0.6 GB |
The download resumes if it is interrupted, and every file is checked against its published checksum before it is used — so a broken download is caught now, rather than blamed on the model later.
It is a one-off. Models are kept outside the program folder, so a later version of Blucta AI will not make you download them again. Settings → Servers → Options shows the folder and lets you move it; an external drive is fine.
Why one big model
Blucta AI ships with one model that drives it, not a menu of them. The smaller ones were tried on the same tasks and measured: one took eleven steps to open Notepad, maximise it, and close it, whereas the 27B took four. Another invented a file of contacts it had never read, and reported success. A model that fabricates is worse than no agent at all, so they are not offered.
That one model does everything, including working out the exact point to click. If you would rather split the job, Settings → Models → Brain can hand the pointing to a second, smaller model instead. It is not downloaded unless you ask for it.
Why 24 GB of memory
It is your computer’s RAM, and memory on the graphics card does not count towards it. The two are used at the same time, not instead of each other: with the model loaded and sitting idle, Blucta AI holds about 19 GB of system memory and about 19 GB on the graphics card.
So a PC with 16 GB of RAM and a 16 GB card is not enough, even though the two numbers add up to more than 24. Blucta AI will still start and still work, but the model no longer fits in memory, and Windows reads it back from disk as it goes.
Pictures and video
These are extra, and nothing is downloaded until the first time you ask for one. Models your card cannot run are not offered at all.
| Model | Makes | Graphics memory | Size |
|---|---|---|---|
| FLUX.2 klein 4B | Pictures and edits of your own | 8 GB | 5.3 GB |
| FLUX.2 klein 9B | The same, better | 12 GB | 11.0 GB |
| WAN 2.2 TI2V 5B | Video, silent | 16 GB | 10.9 GB |
| LTX-2.5 22B | Video with sound | 16 GB | 26.7 GB |
Size and length live in Settings → Models → Media Generator, and both cost time and memory. A short clip at the smallest size takes a few minutes on a fast card; the longest at the largest size takes many times that. Blucta AI knows the trade — tell it a video came out too short or too rough, and it will point you to the setting and say what the better one costs.
Finished work is saved to Pictures\Blucta and Videos\Bluctaand shows up in the chat the moment it is ready.
Optional: using a different model
Everything above works out of the box, and nothing in this section is needed. Blucta AI only ever uses the model you have selected in Settings → Models → Brain → Provider, and that starts on Internal — its own model, on your PC.
If you already run LM Studio, Ollama, or vLLM, you can point Blucta AI at one of those instead. The address is filled in for you and the model list is read from the server, so there is nothing to type. Everything still stays on your machine.
The same list also offers OpenAI, Anthropic, and Gemini, for people who would rather use a service they already pay for. This is entirely opt-in: you have to open Settings, pick one, and paste an API key before anything happens, and you can switch back to Internal at any time. It is also the one arrangement where the screen Blucta AI is looking at is sent off this computer — so if that matters to you, simply leave the setting where it is. Keys are encrypted and stored locally, and are never sent anywhere except to the service you chose.
Whatever you choose, the model must be able to see images. A text-only model cannot drive a screen, and Blucta AI will tell you so rather than fail halfway through a job.
Ports
When it runs models itself, Blucta AI starts them as local servers on 8090, 8091, 8092 and 8093. They listen on this machine only. If something else already has one, change it in Settings → Servers → Options.
Hearing and voice
Voice typing works out of the box. Understanding what your PC is playing — music, a sound effect, someone talking in a video — needs one more model, and you choose when it loads: never, only while something is listening, on first use, or at startup. First use is the default, and it is the right one for almost everyone.
Teaching it an app
A skill is a short file describing how one program is laid out, and how its jobs are normally done. With one, Blucta AI stops rediscovering the same menus every time and goes straight to work. Skills live in their own folder — Settings → Servers → Options — and you can write your own.
When something goes wrong
- “The model server is not responding.” It is still loading. The first start after the PC boots takes a minute.
- Everything is slow. The model is running partly on the processor because it does not fit on the graphics card.
- Pictures or video are not offered. There is not enough graphics memory for any model in that group.
- It clicks the wrong thing. Say so — it looks again. If one program gets it wrong again and again, that program wants a skill.
- It is waiting on a long job. Let it. It checks by itself, and asks you only if it genuinely cannot tell whether the job is still running.
- You want it to stop. Stop it. It drops the current step and hands back to you.