FAQ

General

Do I need internet?

Only to browse the model catalog, download models and engine, and for email. After that, the chat and agents work offline.

Are my conversations sent anywhere?

No. The model runs on your computer and the history is saved there. See Privacy and security.

Which language does the model answer in?

In the language you write in, whatever the interface language.

What's the difference between the chat and an agent?

There is one chat, and it picks the role for each request by itself, using the access you gave it. An agent is a saved assistant with its own instructions, folders, permissions and tools, and with its own conversations. For everyday work the chat is enough; agents are for recurring tasks or ones to set up precisely.

Does the chat remember past conversations?

Only if you turn on memory in Settings. Then it saves short notes about you and searches past conversations when needed; you can see and delete the notes at any time.

Models and speed

The model doesn't start

The chat shows the error and, under Technical details, what the engine said. Usually there isn't enough memory: try a smaller model or a shorter context. The Choose another model button leads to the Models page.

Replies are slow

• With an NVIDIA card, download the CUDA engine.
• Use a smaller model, or shorten the context.
• For simple questions turn reasoning off or shorten it.

The Models page says my RAM runs below its speed

Memory modules start at the standard speed (4800 MT/s for DDR5) until their profile is turned on in the BIOS: EXPO on AMD boards, XMP on Intel. HotMoE recognises this from the module part number (Corsair, G.Skill, Kingston Fury, Crucial Pro) but cannot turn it on: RAM speed and timings are set by the firmware at boot. It matters a lot for models larger than your VRAM, because experts are computed on the CPU reading them from RAM. To turn it on: enter the BIOS at startup (usually Del or F2), enable the EXPO/XMP profile, save and restart. The first boot may stay on a black screen for a few minutes while the board trains the memory: that is normal. With current BIOS versions it is safe; an hour of memory testing is still a good idea (TestMem5, OCCT or MemTest86). If the PC is no longer stable, turn the profile off again.

The reply stopped by itself

If the model was repeating the same text in a loop, the app stopped it: rephrase the question. If it reached the length limit, ask it to continue or to be more concise.

The model reasons for too long

Lower the Maximum reasoning length in Settings: when it runs out, the model closes its reasoning and answers.

The chat says the conversation was compacted

The conversation was close to the model's context limit, so the app summarized the oldest messages. The history still has them all; only the model sees the summary, which can be less precise. If you need an old detail, repeat it. A bigger context means fewer compactions. You can also compact yourself with /compact, adding what to keep. See Long conversations.

"The request does not fit in the model's context"

Even after compacting the conversation, the request is too big, for example a single huge message. Start a new chat or increase the context in Settings.

Can the chat read images?

Yes, with a model that sees images: on the Models page it is marked Sees images. Attach the image with the button in the message box, drag it onto the chat or paste it with Ctrl+V. See Images.

The Models page says Hugging Face is not responding

The page shows the last catalog saved on your computer. The models already on disk start as usual; to download a new one, Hugging Face has to be reachable again.

The working chat and agents

The chat doesn't look at my files

Check that you granted the folder under Access. If the chat doesn't activate the right role, add it yourself with + role: the next message will use it.

The agent says it's done, but it didn't do everything

It happens with small models. Ask it to check ("make sure all the photos were moved") or use a model marked Good for agents.

I allowed an action by mistake

Open the Log (in the chat, next to the history; in agents, on the right) and use Undo on the action. Commands that ran and emails that were sent can't be undone.

How do I remove an "Always allow"?

In the chat: Access › Permissions › Revoke. For an agent: Edit, Permissions section.

Email

Gmail rejects the password

Gmail doesn't accept your normal password: create an app password (2-Step Verification is required) and use that. The link to the right page appears in the form when you type a Gmail address.

Can I use Outlook.com or Hotmail?

Not yet: Microsoft only accepts its own sign-in, which will come in a future version.

Will the app send emails on its own?

No. Every send shows you the full message, which you can correct, and waits for your yes, and it can't be allowed forever.

Data

How do I start over?

To redo the first launch, close the app and delete settings.json in the data folder. To remove everything, delete the whole folder. See Where data is.

Can I move an agent to another computer?

Yes: each agent is a file in the agents folder (Open the agents folder in the editor). Granted folders and email accounts then need to be linked again on the new computer.

HotMoE is a Virsion project. This guide describes the app as it is today: what is still on the way is marked as such.