v0.10.3

Pumpkin 0.10.3

New

Teammate has its own page in the left menu. Pumpkin works in your own Chrome or Edge, with your logins, and asks before it uses a site.
The extension setup (download, pairing code, allowed and blocked sites) moved there from Devices.
Teammate keeps a list of past browser chats. Open one to see what it did step by step, or delete it.

Changed

The pairing text in the extension and the guide now point to the Teammate page.
The Teammate page doesn’t show on Android.

Full Changelog: v0.10.2...v0.10.3

v0.10.2

Full Changelog: v0.10.1...v0.10.2

v0.10.1

Pumpkin 0.10.1

• The update now reliably starts the new version. If the old Pumpkin was slow to quit (a loaded model takes a while to unload), the update could finish with nothing running and you had to start Pumpkin by hand. It now waits for the old one to stop, then starts the new one.
• The browser extension panel shows one screen at a time. Before, the pairing form and the chat were both visible and the message box was out of view.
• The update log on the Settings page no longer shows colour codes as text.

Full Changelog: v0.10.0...v0.10.1

v0.10.0

Pumpkin 0.10.0

Helper agents. Pumpkin Code can hand a big reading job to a helper with a fresh context that can only look. It comes back with a short summary, so the main session's memory stays free.

Workspaces. A folder with its own instructions, memory, connectors, skills, model and approval mode. Works for code, writing, data or any files. Reads Word, Excel and PowerPoint. In a terminal: pumpkin code -w NAME.

Pumpkin in your browser. A Chrome/Edge extension that lets the model on your computer read pages, click and type in your own browser.

• Pair it on the Devices page with a one-time code. The paired key only opens the browser part of Pumpkin.
• It asks which sites it may use. Banks, mail, password managers and your own computer or home network always ask and are never kept.
• It asks before each action if you want it to, and shows what a button really says.
• It never types passwords or card details, and Pumpkin's own page is off limits.
• Page text is treated as data. If a page tries to give the AI orders, every action asks until you write again.
• A browser chat can't touch your files or run commands.

Full Changelog: v0.9.1...v0.10.0

v0.9.1

Pumpkin 0.9.1

Fixes

• GPU code that won't compile. Some new llama.cpp builds can't compile their GPU code on an older macOS, so no model starts. Pumpkin now spots this the first time a model fails to start, goes back to the build before, and skips that build and newer ones for a week.
• Honest error message. The error says the build can't use the GPU. Before, it blamed memory. Pumpkin also stops halving the context for nothing in this case.
• Model server card. It warns when the build in use has this problem.

Full Changelog: v0.9.0...v0.9.1

v0.9.0

Pumpkin 0.9.0

llama.cpp stays up to date

• Update button. The Model server card has an "Update llama.cpp" button.
• Your choice. Pick Off, Tell me, or By itself. Pick the latest release (a day old) or the newest build.
• New builds get tested. Pumpkin checks a new build the first time it runs. If the chat breaks, it goes back to the previous build and skips the bad one.
• Go back. The button is always there. Your own builds are never replaced.
• Older systems. If a build won't run on your system, Pumpkin skips newer builds for 30 days.
• Setup. It installs the newest llama.cpp and falls back to the pinned b11796 if needed.
• Update and try again. On the Overview, it uses the newest build so brand-new models load.

Models and presets

• Vision files. Presets show clearer file lists with folder labels and warn about mismatches. The mmproj file switches automatically when you change the model.
• Speculative decoding. It's now a main toggle in presets. The measured result shows under Reports → Speed.
• Compare tab. Reports has a new tab for comparing two models.

Overview and pairing

• Memory notice (Linux only). The Overview warns when the model is swapped out or reading itself from disk.
• Pairing. The model restarts itself after you pair a device, once it's idle.

Full Changelog: v0.8.8...v0.9.0

v0.8.8

Pumpkin 0.8.8

•	Pictures on older Macs. Pumpkin now picks a stable-diffusion.cpp build that runs on your macOS. It used to download the newest one, which is built for macOS 26 and crashes on Sonoma and Sequoia with "The image engine stopped". If the newest doesn't run, it tries older releases, and only tells you to update macOS if none work. If you tried it before and it failed, delete ~/.local/share/pumpkin/sd so it downloads again.
•	Under the hood: the build tools used by the GitHub checks and releases are updated, and Android build tool upgrades are no longer proposed automatically.

What's Changed

• Bump the actions group across 1 directory with 9 updates by @dependabot[bot] in #1

New Contributors

• @dependabot[bot] made their first contribution in #1

Full Changelog: v0.8.7...v0.8.8

v0.8.7

Pumpkin 0.8.7

•	The Update button now reloads the page as soon as Pumpkin is back after restarting, so it no longer sits there spinning.
•	While it updates, the button shows what the installer is doing ("Downloading", "Installing") instead of just a spinner.
•	If an update takes longer than 4 minutes, it still shows the manual command and the log.

Full Changelog: v0.8.6...v0.8.7

v0.8.6

Pumpkin 0.8.6.

Fixed the “Update available, reload” popup that showed up on the first start after every install or update, even though the page was already the newest version. It now only appears when the page you have open is actually older than the server.

Full Changelog: v0.8.5...v0.8.6

v0.8.5

Pumpkin 0.8.5

Two fixes.

•	Photos in a chat with a text-only model. Pumpkin reads attached photos with a helper vision model. If the chat model and the vision model didn't fit in memory together, the chat model was stopped and restarted once per photo, so a few photos looked like an endless load/unload loop. Now all the photos in a message are read in one go, so the chat model swaps out and back once. If the vision model fails to start, Pumpkin tells the chat model the picture couldn't be read and stops retrying for 2 minutes, instead of swapping on every message.
•	Clearer error when a model won't load. If llama-server doesn't know a model's architecture, Pumpkin now says your llama-server is too old and names the architecture, and tells you to update it in Settings or pull and rebuild it if you built it yourself. Before, you got a generic "file may be incomplete" message.

Full Changelog: v0.8.4...v0.8.5