GIDE 2.4.2: Model Hub, VPS & Private Servers, and images on local models

GIDE 2.4.2 takes the guesswork out of getting a model running, wherever it runs. A new Model Hub finds and tests a local model that fits your machine. VPS & Private Servers finishes what 2.4.0 started, with a guided setup for people who don’t already run a model server. And images now work with local models, not just online ones.

At a glance

• Model Hub. Search Hugging Face or pick a recommended model, and GIDE downloads the version that fits your memory and tests it before use.
• VPS & Private Servers. A model on a VPS, VPN or tailnet you control becomes a first-class row in the model picker.
• Guided server setup. Answer two questions, run one command on the server, and GIDE tells you exactly what to enter.
• Local models read images. A local model with a vision file reads images directly.
• Online image support and prompt caching. GIDE learns which online models can read images, and long online conversations stop resending the same history.
• Check for Updates works everywhere. On macOS, Windows and Linux.

Model Hub

What changed

Getting a local model used to mean finding a GGUF file yourself, picking a quant and hoping it fit. A new Models view in the sidebar does that for you. You can also reach it from Browse models… in the model picker.

Watch the Model Hub in action:

How it works

Pick one of 14 recommended models or search Hugging Face. Trusted publishers are shown by default. GIDE picks the largest Q4_K_M or better version that fits your memory, plus the vision file that goes with it.

Before a model is marked ready, GIDE:

1. checks the download against its SHA-256,
2. loads the model,
3. has it answer one prompt and read a test image.

Downloads resume after an interruption and fall back to GIDE’s own mirror if Hugging Face fails or is blocked. They use your system’s certificates and proxy settings, so they work behind company networks.

The Installed tab lists every model on this computer. Delete is only offered for models GIDE downloaded, never for a file you added yourself.

From the CLI

gide models search <query>
gide models pull <model>
gide models list

VPS & Private Servers

What changed

2.4.0 added the option of using a model on your own server. In 2.4.2 that support is complete. The model picker has a new row, Your Server. Point it at a model on a VPS, VPN or tailnet you control and use it like any other model, with your prompts staying on that network.

Safe by default

• Plain http is only allowed when the address resolves to your own machine or a private network. GIDE checks again on every connection, so a DNS change can’t redirect your prompts.
• Self-signed certificates are saved on first connect and refused if they change later.
• The key is optional on private addresses. When you give one, GIDE checks it against the chat route itself.
• Web search stays off while a server connection is active, unless you allow it.

The context window is read from the server, and an unreachable server times out with a hint to check your VPN instead of hanging.

gide connect --server
gide servers

Guided server setup

Most people don’t already have a model server running. Open Your Server → Set up a server, or run GIDE: Set Up Your Server from the Command Palette. Answer two questions, and GIDE writes the steps for your computer with your details filled in.

On the server, one command (serve.sh on macOS and Linux, serve.ps1 on Windows) detects the CPU and GPU, downloads a pinned server build and GIDE’s default model, verifies both and starts the server. It listens only on its own machine unless you pass --lan, which also generates an API key. At the end, it prints exactly what to enter in GIDE.

The full guide is on the VPS & Private Servers page.

Local models read images

In 2.4.0, images only worked with online models. A local model with a vision file now reads them directly, and models from the Model Hub come with theirs.

• Attached images show as thumbnail tiles.
• Pasting an image file attaches the image instead of its path.
• You can send up to 4 images per message.
• Every Mac image format works. WebP, AVIF, BMP, ICO and SVG are converted inside GIDE, and HEIC, HEIF and TIFF (an iPhone photo, for example) by the local server.

Images from your last session no longer come back after a restart.

Online models

Knowing which models can read images

About a third of OpenRouter’s models can’t read images, so GIDE no longer assumes they can. Each connection finds out from the provider’s model list, from your own server, or with a quick one-time test at connect, and it remembers when a provider rejects an image. Once GIDE knows a model can’t read images, it tells you before you send one, and text follow-ups keep working.

GIDE only ever learns “no” this way, never “yes”, because some models accept an image without an error and then answer without looking at it.

Prompt caching

Every online connection now uses prompt caching, so a long conversation doesn’t pay to send the same history again on every step. Anthropic, OpenAI and Claude models behind other gateways get explicit cache hints, and other providers cache automatically.

Online models now get the conversation in its original order. In our tests, that took the share of each agent step read from cache from 42% to 94-100%. The system prompt is also sent once instead of up to three times.

Online model fixes

• OpenAI’s o-series reasoning models work in the agent. They reject max_tokens, so GIDE now sends the field each provider expects.
• A model’s real output limit is learned from the provider’s error, and the turn is retried.
• Anthropic models get their real context window (200K for Haiku instead of 32K), and any provider whose model list states a window is read the same way.
• An unrelated error no longer switches tools off, and the first word of an answer is no longer lost.
• Conversation titles are titles instead of raw JSON, and a skipped tool call no longer breaks the conversation with an HTTP 400.
• General questions get general answers again. A word shared with GIDE’s docs, such as “error”, no longer makes “What does HTTP status 404 mean?” count as a question about GIDE.

A labelled chat toolbar

The model button, the settings gear and the tools icon used to be icons with tooltips, and new users couldn’t tell what they did. The model button now says Choose model until one is selected, then shows its name. Model settings and Tools are labelled with words, with Model settings beside the model button as one split button. A narrow chat panel shortens the labels, then shows icons only.

Updates

Check for Updates

The Check for Updates button asked an update server GIDE doesn’t use, so every manual check ended in an error on macOS, Windows and Linux. The button, the menu, the Command Palette and the launch check now all use GIDE’s own updater. You get Update Now or Later when you’re behind, “GIDE is up to date” when you’re not, and an explanation when you’re offline.

Linux tarball installs are told when a newer version exists, and update.mode and --disable-updates are respected.

Windows updates

Windows hadn’t offered updates since 2.3.4. It now uses the same consent dialog and progress window as macOS: GIDE downloads the installer, checks its SHA-512, installs silently once GIDE closes, reusing your install folder, and reopens. If you choose Later, the verified installer is kept for next time.

Other fixes

• With two accounts on one Mac, the app, the workbench and the CLI could attach to the other account’s GIDE server. They now use a server only if it accepts your own sign-in token, and Show Hardware Info asks the right server too.
• Searching the codebase with no project open now says so.
• Tools and Model settings keep a fixed order.
• The Models view stops spinning when GIDE’s server is down.
• The chat-only online path respects a custom endpoint’s URL.

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