GIDE 2.4.0: two models at once, @ mentions, your own VPS and documents

GIDE 2.4.0 is about models working side by side instead of one at a time. Until this release, GIDE loaded exactly one local model, and local and online models were either-or. Now a second local model can stay loaded, any connected model is an @ away, and a model on your own server joins the list.

The release also brings images and documents into the conversation, a sign-in flow that browsers can’t block, and a CLI that checks your machine before it recommends a download.

At a glance

• Two models at once. Keep a second local model loaded when memory allows, and switch between loaded models instantly.
• @ model mentions. Send one message to any connected local or online model without changing your default.
• VPS & Private Servers. Connect to a model running on a virtual private server you control.
• Images and documents. Attach images, plus Word, PDF, Excel, PowerPoint and CSV files.
• Sign in with a link. Login comes back to GIDE through a gide:// link.
• Update progress and CLI improvements. Measured progress while updating, a CLI first run that checks your hardware, and new model commands.

Two local models at once

What changed

Before 2.4.0, GIDE loaded exactly one local model, however much memory your machine had. Switching models meant restarting the model server and reloading several gigabytes of weights.

GIDE now keeps a second local model loaded when the remaining memory genuinely fits it. The model you pick answers, and the other one stays a mention away. Switching between loaded models is instant: it’s a pointer move, not a restart, and an attached CLI session keeps its warm state.

Nothing shrinks silently

The model picker shows free memory above the list and marks any model that won’t fit alongside the current one, with the reason. Keeping two models loaded can cost one of them part of its context window. When it does, the row tells you before you choose it, for example:

both loaded: 32K → 16K

A loaded model that isn’t answering has an unload button, which frees its memory without a restart. On machines at the hardware floor nothing changes: one model gets the whole memory budget.

From the CLI

# Load a second model without unloading the current one
gide models keep <model>

# Free one loaded model and leave the rest
gide models unload <model>

gide models keep warns you first if keeping both would shrink a context window.

More accurate memory sizing

The memory estimate behind model sizing was rebuilt. Instead of computing a figure, GIDE now simulates the runtime’s allocator for the exact model file. On every model and context size we checked, that matched the runtime’s own measurements. Single-model sizing gets more accurate too, even if you never load a second model.

@ mention any model

What changed

Local and online models are no longer either-or. Connect a local model, several online models, or both kinds together, in any order. Picking a local model no longer disconnects your online ones.

How it works

Type @ and pick a model. Every model that can answer right now is listed first, marked loaded or connected. The mention routes that one message without changing your default, and the model you address sees the conversation so far.

Replies say who answered on the first reply, when the answering model changes, or when you used @. When the same model answers again, they stay quiet.

Each connection keeps its own key in your OS keychain, and GIDE reconnects them when the server comes back. File, symbol and terminal mentions such as @src/app.ts or @errors work exactly as before.

Why it matters

You can bring a different model into the thread for one question, with the full conversation behind it, and then carry on with your default. You don’t have to switch, re-explain the task or start a new session.

VPS & Private Servers

GIDE can now connect to a model running on your own virtual private server (VPS). Keep the model on a machine you control, reach it privately, and use it from GIDE like any other model, with your prompts staying inside your own network.

This was the first step. GIDE 2.4.2 completed the support with the Your Server row in the model picker and a guided server setup.

Images and documents

Images

Paste an image, or attach one with +, and ask about it when the model you’re using can read images. GIDE reads whether a model takes images from what its provider reports, so a new vision model works without a GIDE update. A model that can’t read images says so plainly instead of the image being dropped.

In 2.4.0, images work with online models that support them. Local models are text-only for now. A file the model can’t read, such as a binary, is no longer sent to it as raw bytes; the attachment is marked “not readable” instead.

Documents

Attach a Word, PDF, Excel, PowerPoint or CSV file and ask about it. This works with local models, online models on your own key, and Private Cloud. Drag a file in, or use + → Upload from computer.

A document stays attached for the whole conversation. Answers cite their source, such as page 4, Sheet1 or slide 12, and each format gets its own icon.

How much of a document the model reads depends on the model:

• An online or Private Cloud model with a large window can read a long document whole, about 200 pages on Private Cloud.
• A local 32K model keeps the same budget as before.
• When a document is larger than the model can read, GIDE searches it for the parts that match your question.

Questions that need the whole document, such as counting or totalling, are refused rather than answered from a partial read.

Limits: 20 MB per file, 10 files per message, and a separate cap on Word files.

Sign in with a link

Some people couldn’t sign in at all, depending on their default browser. Login used to finish with the website calling a server on your own machine, which many browsers and antivirus tools now block.

Login now returns to GIDE as a gide:// link, which no browser treats as a network request, and signs you in to the app, the CLI and the server together. The exchange is protected with PKCE, with a visible code as a fallback.

gide login

In the CLI, gide login uses the same flow and ends with you pasting the code, so it also works over SSH. Pressing Send while signed out now shows a sign-in link instead of opening your browser, and the Login to GIDE links in chat work.

Smaller improvements

The model button shows the model name

The chat input’s model button now shows the name of the model you’re talking to, instead of an icon with a “Pick Model” tooltip. Model settings moved onto the model itself: the loaded local model and every connected online model carry their own settings gear in the picker, and a local model is shown with a chip icon.

Update progress you can see

After you agreed to an update, GIDE used to go silent for minutes. A progress window now shows each step with measured progress: bytes received and transfer speed while downloading, bytes checked while verifying, and bytes written while installing.

Progress also shows on the Dock or taskbar icon, and you can cancel the download. A download that ends short is rejected as incomplete, and a failed download removes its partial file.

A CLI first run that fits your machine

The CLI now runs the same hardware check as the app before offering a download, and recommends a model that suits your machine instead of a 6 GB download on every machine. Every local model is on offer, each saying what it needs. “Back” now goes back instead of quitting (only q quits), and an installed model is recognized instead of the CLI asking you to find its file.

To remove the model you’re using, open /model and choose Remove a model…. GIDE disconnects it first, and deleting the file is only offered for models GIDE downloaded itself.

CLI terminal polish

• Typing works straight after a model switch on Windows.
• Keys typed while the input box is appearing are kept.
• Windows gets a proper file picker with modern styling and DPI awareness.
• The spinner and thought markers were redrawn.
• Pressing Enter on an empty input no longer sends a blank message.

Fixes

• Attachments and mentions. Online and Private Cloud models ignored every attachment, code files included. Questions about an attached document were sometimes answered from GIDE’s own manual, and documents were uploaded again on every turn. OpenRouter models, whose names contain a /, couldn’t be mentioned, and mentioning a model with spaces in its name sent part of the name as your message.
• Online connections and the model picker. Disconnecting a local model ended every online connection for good. The picker could show an online connection that no longer existed after a server restart, deleting one online model disconnected a different one, deleting a built-in provider removed an unrelated saved endpoint, and local model settings were unreachable while any online model was connected.
• Login and packaging. Saving a login session never worked, a failed file write was reported as a bad login code, and opening a gide:// link in development started a second copy of GIDE. Packaging now refuses to ship a server with a missing dependency, since a single missing package stopped the whole model server from starting.

For the short version of every change, see the GIDE changelog.