v0.2.1-nightly

noullama 0.2.1 nightly

A pre-release for testing in-app updates. Packages: .deb and .rpm for 64-bit Linux.

In-app updates

• Updates in the sidebar shows your installed version and checks GitHub for a newer release. When one exists, the entry changes to Update to x.y.z.
• Install and restart downloads the package matching your install (.deb or .rpm), checks its signature, and asks for your password through the system prompt. noullama then unloads the model, stops the local API, and restarts on the new version.
• Check when noullama starts checks once shortly after launch. Nothing downloads until you choose to install.
• Include pre-releases also offers nightly and pre-release builds. It starts switched on for this build because it's a nightly.
• Copies built from source, or run as a standalone executable, still show when a new version exists but must be updated by hand.

Upgrading to this release

Versions before 0.2.1 have no updater. Install this release once by hand. Future updates can then be installed from inside the app.

• Fedora and other RPM distributions: sudo dnf install ./noullama-0.2.1-1.x86_64.rpm
• Debian and Ubuntu: sudo apt install ./noullama_0.2.1_amd64.deb

Other changes

• Release packages are now signed, and each release includes an update manifest (latest.json) for the updater.
• Release builds now run on Node.js 24, which clears GitHub's Node 20 deprecation warnings. Building from source now needs Node.js 24 or later.
• README: new Updates section and the Node.js 24 requirement.

Known issues

• If you cancel the system password prompt, a second password box may appear. Cancel that one too to stop the update.
• The updater installs the package file directly, without dnf's or apt's dependency resolution. If a future release adds a new system library, update with your package manager instead.

v0.2.0

Noullama 0.2.0

A redesigned interface, image input for vision models, and clear errors when a model fails to load.

New look

• Sidebar navigation, with a runtime card at the bottom showing whether a model is ready.
• Page headers show this machine's graphics memory, system memory and CPU threads.
• Playground: the state and results sit on the left, with presets and the questions editor in a Run settings panel on the right. Long JSON lines now wrap.
• Hub: search results on the left, the picked repository's files on the right. Type organization/repository to open a repository directly.
• Downloads card: a card in the bottom-right corner shows download progress anywhere in the app, with a Cancel button.
• Runtime tab: regrouped into llama.cpp (installed builds, custom path, downloads), Model and Connection. The save bar floats only while you have unsaved changes.
• Dropdowns, the right-click menu, number fields, switches and the editor's fold markers now match the app instead of using system controls. Hold Shift while right-clicking for the system menu.

Images

• A Vision switch loads a model's multimodal projector (mmproj) when one sits in the same folder. Projectors are no longer listed as separate models.
• With a projector loaded, add images in the Playground by button, drag and drop, or paste. You can leave the text empty to ask about the images alone.
• WebP and other formats are converted to PNG automatically. Each image can be up to 20 MB, and all images together up to 60 MB.
• The local API accepts images on /v1/systemone as an images array of base64 data URLs.

Clearer errors

• When a model fails to load, or the runtime stops on its own, a red panel under the model bar shows the reason and the error lines llama.cpp printed.
• Sending images to a model loaded without a projector now tells you to turn on Vision and reload, instead of showing a bare 501 error.

Model details

• Hover the runtime card for the loaded model's device, projector, context size, GPU layers and API address, plus live graphics memory: the model's own use on NVIDIA and total card use on NVIDIA and AMD.
• The memory-fit estimate now includes the projector's size.

Fixes

• Dropping a file outside the State box no longer replaces the app window with that file.
• A cleared number field no longer snaps to 0 while you retype it.
• Text pasted together with an image is no longer dropped.

Linux .deb and .rpm packages are attached below and linked from noullama.dev/download.

v0.1.2

Noullama v0.1.2

Playground

• Proper JSON syntax highlighting. Each kind of value in the questions and state editors now has its own color: keys blue, strings green, numbers orange, true, false and null red, punctuation grey
• Visible text selection. Click-and-drag now shows a clear highlight in the editor, and matching brackets are highlighted too

Downloads

• Cancel works right away. Cancelling while a download is still preparing now stops it instead of being ignored
• Partial files are actually cleaned up. After you cancel a download, its partial data is now deleted when you quit Noullama. Before, the largest partial file was missed and left behind

Runtime

• Slow connections are fine. llama.cpp downloads no longer fail after 15 minutes. They only stop if no data arrives for 60 seconds
• Large models can load. Model loading no longer gives up after 90 seconds. It fails only if the runtime goes silent for 5 minutes, which leaves room for big models and first-run CUDA setup
• Interrupted installs are cleaned up. Leftovers from a runtime install that was interrupted at quit are removed on the next launch

App

• Installed and source builds coexist. If an installed package and a build from source are both launched, the newer one replaces the older instead of failing to start
• Faster first load. Hardware detection runs in the background, so the window no longer waits on nvidia-smi

Install

Download the .deb or .rpm for 64-bit Linux from noullama.dev/download or the assets below. This is an early beta, so please report bugs on the issue tracker.

v0.1.1

Noullama v0.1.1

Downloads

• Two-sided progress bar. The download panel now shows network receipt and file reconstruction as two bars. They start at opposite ends and fill toward the middle, so you can see both stages at once.
• Cancel button. You can now cancel a model download while it's running. The partial files are kept so an immediate retry resumes, and they're deleted when you quit Noullama.

Hub and Models

• Choosing a repository in the Hub now scrolls straight to its file list and download section.
• The Models tab rescans your model folder every 5 seconds, so the list no longer shows outdated files.

Documentation

• The README is reorganised with clearer sections and subsections. The content is unchanged.

Install

Download the .deb or .rpm for 64-bit Linux from noullama.dev/download or the assets below. This is an early beta, so please report bugs on the issue tracker.

v0.1.0

First Official Release of Noullama