1.26307 — 3 November 2026

Until now, gezel could draw a new picture but could not change one you already had. Now it can, and any picture can have a style. You can also measure how well a model does real work on your own computer, with the same tests we use for the published scorecard. New Project has nineteen ready-made projects for running a small business. And gezel now comes as a Docker image too, so a server or a computer without a screen can run all of it, reached from a browser.

If you read only one thing: add the Image editor, attach a photo, and say what to change.

Change a picture you already have

Asking for a picture used to mean starting from words, every time. The catalog now has an Image editor, a gezel that changes a picture you give it. Attach a picture and say what to change, such as "remove the background" or "turn it into a watercolor." Your message goes straight to an image model that edits, with no chatbot in between, and each follow-up keeps editing the last result. A message without a picture gets a short reply asking for one, not a new picture drawn from your words.

The catalog also adds a model made for this: Qwen Image Edit 2511. It changes what you ask for and keeps the rest. It holds faces steady across several edits, rewrites the text inside a picture, and can combine up to three pictures into one. It is about 19.5 GB to download, and it is slow: one edit takes several minutes. Its fast mode, below, cuts an edit from 20 steps to 4.

Gezel now also keeps track of which of your image models can edit. A request that comes with a picture goes to one that can, and a model made only for editing is never used to draw from words alone. More in Image generator.

Styles and fast mode for pictures

Settings → Image generation has a new list: Styles and fast mode. A style gives every picture the same look. The first ones are pixel art, soft watercolor, ink wash, gouache, paper cut, coloring-book line art, a clean 3D render, an isometric 3D object and a child's drawing. Each works with particular models, and the list says which. Once one is installed for the model you use, the strip above the image gezel's message box gets a Style picker. Choose No style to go back to plain pictures.

Fast mode draws a picture in a few steps instead of dozens: 4 instead of 8 for Krea 2 Turbo, about 1.6 times faster, and 4 instead of 50 for MOSS-Image. Once installed, fast mode is on for its model. To turn it off for one model, use the Fast mode switch in that model's row of the Local models table.

Measure a model on your own computer

The published scorecard tells you how a model did on our computers. Now you can run the same tests on yours. In the terminal, gezel bench runs the scorecard's core set on the model your chats use, and gezel bench --suite all runs all four sets. Settings → Benchmarks runs the same tests and is now there for everyone, where before it appeared only in debug mode. Both start by naming what they measure on: the hardware, the engine build and the gezel version. A job tried fewer than three times shows as a count, never a percentage, just as on the scorecard. In the terminal, a run that could take more than an hour asks before it starts.

To share a result, gezel bench submit <file> sends it to us as a GitHub issue. Results we accept appear on the model scorecard under Community rounds. They stay apart from our own rounds, because we did not control the hardware.

The scorecard itself has six new rounds. Three compare Qwen 3.8 27B with Gemma 4 12B on a different machine each: an NVIDIA DGX Spark on Linux, a Windows computer with an AMD Radeon R9700, and a Windows laptop with NVIDIA's RTX Spark N1X. On the DGX Spark, Qwen passed 106 of 129 attempts and Gemma 90. The fourth round runs Gemma 4 12B through every task set on a Windows laptop with an RTX 5070 Ti. The last two compare the two engines on an Apple M5 Max, MLX and llama.cpp, with Qwen 3.8 27B and Gemma 4 12B on each. In the larger one, 516 attempts across all four task sets, both models did better on llama.cpp: Qwen passed 104 of 129 there against 92 of 128 on MLX, and Gemma 86 of 128 against 67 of 129. A shorter follow-up after fixes ran 18 jobs per model on each engine, and both models passed 16 on MLX and all 18 on llama.cpp. A new table, Every result in one table, lines up every round side by side, and you can download all of it as one spreadsheet. How we test models now lists each test in each set.

gezel doctor now also checks the model it will use. It says whether the model starts, answers a tool call, writes fast enough and has room for a prompt, and how often it finished real jobs on the scorecard, with the hardware named. gezel model list --available shows each model's measured pass rate beside its size.

Run gezel in Docker

Starting with this release, gezel also comes as a container image, for a server or any computer that runs Docker. One command runs the whole of it, with every engine for local models already inside. Besides the models you choose, it downloads a browser for web pages and the small models it uses for search the first time it needs them.

docker run -d --name gezel -p 127.0.0.1:6228:6228 -v gezel:/gezel ghcr.io/bendyline/gezel

The container's log has a link that creates your account, and from then on you sign in with a name and password from your browser. The same container gives your other programs local AI through an OpenAI-style address: make a key with docker exec gezel gezel integrations token create <app>. It can also run one craftbook on a folder and hand back what it made, with gezel do … --out <folder>. There are images for NVIDIA cards (:cuda) and for AMD and Intel cards (:vulkan), on GitHub's registry and on Docker Hub as bendyline/gezel. A new image comes out with every release, tagged with the same version as the gezel command-line tool. Craftbook steps that run small scripts with no network access run inside the container too, in a network sandbox of their own.

Everyone you add in Settings → Accounts can do everything you can, so add only people you trust. More in Gezel in Docker.

Nineteen new projects for running a small business

New Project has nineteen new project types for the everyday work of a small business. Each comes with its own crew and a page that keeps the work in view. For money: Books & Month Close, which books every receipt into a categorized ledger and works through a month-close checklist with you, Cash Flow Planner, which keeps a 13-week forecast from the invoices and bills you expect, and Grants & Proposals. For customers: Quotes & Bids, Support Desk, Customer Voice and Client Delivery. For your team: Hiring Desk, New Hire Onboarding, Shift Roster, Training Tracker and Team Goals. And for the rest of the business: Campaign Room, Content Studio, Event Desk, Inventory & Reorder, Vendor & Contract Desk, Compliance Register and SOP Library. A twentieth, Sales Pipeline, appears once you turn on Show preview features in Settings → About, because one of its craftbooks can read your calendar through a connector, and connectors are still a preview.

They bring eleven new crew roles, among them the Bookkeeper, Recruiter, Estimator and Stock Keeper, and six craftbooks for jobs that come back every week or month: a reorder run that turns a stock check into purchase orders, replies to customer reviews in your own voice, a structured interview kit for a role, a week's shift schedule, a run of show for an event, and an audit evidence pack from the compliance register.

The craftbook editor's AI assist stays on the craftbook

The AI assist in the craftbook editor is for improving a craftbook. It used to have the full set of tools, and it sometimes ran the craftbook or moved on to other work instead of editing it. Now it has only the tools for editing craftbooks, looking up other craftbooks and the crew, reading the Handboek, and asking you questions.

Each message to it now starts with the craftbook's current outline and a short checkup. The checkup flags a step with no check that proves its work, a step with no role assigned, an input that asks you to type a file path, and a missing description or trigger phrases. So suggestions are based on the craftbook as it stands, and the checkup shows when a change has fixed a problem.

Overview and AI assist now sit in the same row of tabs as the steps. When you go back to Chat, it no longer reopens a craftbook conversation. When you edit one of a craftbook's scripts, Back to craftbook asks first if you have unsaved changes.

Privacy and security

None of these need anything from you.

Camera pictures. The last release noted one exception: asking a cloud gezel or a cloud image engine to make a new image from a camera moment's pictures still sent those pictures. That gap is closed. Unless you let cloud AI see a camera's pictures, a cloud image engine is refused, and the message says to pick an on-device model or change that camera's setting. More in Cameras and lookouts.
No usage reports from ONNX Runtime. ONNX Runtime is the library gezel uses for search, face and photo recognition, and speech. On a Mac and on Linux it queued usage reports for Microsoft that named the computer model, the operating system and a device id. Gezel now switches that off before the library starts. On Windows it follows your Windows diagnostic-data setting, as before.
Updated components. The MCP SDK, undici, Hono and nodemailer are on newer versions, and our release checks now stop a release if any published package has a high or critical security advisory.

New in the Handboek

What leaves this computer lists every kind of network traffic gezel can make, what it carries, and which security level allows it.
Governance explains who writes gezel, how an idea from outside reaches a release, and what keeps working if we stop.
Gezel on a server covers running gezel without a screen, on a computer of its own, with the security level and activity mode a server should have.
Gezel in Docker covers the container: the workshop in a browser, keys for your other programs, folder jobs, graphics cards, accounts and updates.

Smaller things

Windows updates land. Before installing an update, Windows checks its signature, and on a 1 GB installer that check could take longer than the 20 seconds it was given. The download was then thrown away and fetched again at the next launch, so some updates never installed. The check now has time to finish.
A step that needs to run tests says so straight away. Some craftbook steps prove their work by running the project's own tests, such as npm run test. While Run package scripts is off, which it is below Unrestricted, no gezel can run them, and such a task used to keep retrying for up to an hour. Now it pauses at once and names the switch in Settings → Security & Compliance that would let it go on.
Clearer pause messages. A paused task's card now gives the reason in one plain sentence, such as "The work stopped getting closer to passing its checks," instead of quoting the check that failed.
No false "started" message. When the Meester starts a craftbook whose first step has no one assigned, the task waits for you, and the reply now says so instead of reporting the work as underway.
The request card reads properly. It used to say things like "Soren is acquiring and verify sources." Now it says "acquiring and verifying sources."
Pictures for a task. When a task step goes to the image gezel, the picture now matches the brief, not the wording of the handoff message.
Tasks stay in view when Windows locks a file. If another program, such as a sync or antivirus program, holds a file for a moment, active work no longer drops out of view.
Project settings are grouped. A project's Settings tab sorts its settings into Workspace, Access, Toolsets, Preferences and History.
Files, not References. In the chat rail, the tab that lists the files a conversation used or made is now called Files.
One fewer badge in the titlebar. Background indexing now shows in Activity with everything else, not in a badge of its own.
DocBlocks is no longer downloaded again for a project when the copy bundled with gezel is as new or newer.
The setting that turns on early features is now called Show preview features.
New in the catalog. Wikipedia: General Knowledge covers concepts, organizations, awards, events and schools that no subject catalog covers, and the Wikipedia subject and people catalogs are refreshed.

In the terminal

gezel bench runs the scorecard's tests on this computer. gezel bench list shows every set and how long it can take, gezel bench results shows finished runs, and gezel bench submit <file> shares one.
gezel security level shows the security level and what it allows. gezel security level set super-lockdown|lockdown|free switches to a preset. free is called Unrestricted in Settings.
gezel doctor checks whether the model it will use can do the work, as described above.
gezel model list --available shows each model's measured pass rate.
--no-start on gezel craftbook list and gezel model list answers only from a service that is already running and never starts one.
gezel account adds, disables, removes and resets the passwords of the people who sign in from a browser. gezel start --web --accounts uses those sign-ins instead of a one-time link.
gezel integrations token create <app> makes a key for another program without the approval prompt; revoke takes it back.
gezel do … --out <folder> waits for the task and copies what it made into that folder.
The reference now shows how to install Node.js 24 on macOS, Windows and Linux.

See the CLI reference.

For developers

Pages. gezel.ui.components.chat(element, options) puts gezel's own chat inside a project-type page, and can lock it to one role from the project's crew. gezel.ui.layout and gezel.ui.onLayout() tell a page whether it fills the window, so it can lay itself out differently when maximized. See Page APIs.
App SDK. The images client lists the catalog's styles and installs one through /v1/images/styles and /v1/images/styles/ensure. Callers name a style by id and never pass a path, URL or engine argument.
ONNX telemetry. The service, inference and knowledge packages set ORT_DISABLE_TELEMETRY=1 before ONNX Runtime first loads, unless you have set it yourself.
Containers and accounts. With GEZEL_CONTAINER=1 inside a container, the service adds a listener on the container's network that admits only signed-in accounts and app keys, and leaves out the routes that trust a local caller. Accounts live in accounts.json and account-sessions.json, apart from tokens.json; History records which account did what. A home shared by two containers now refuses the second, and memory and thread counts follow the container's limits. See docs/container.md and docs/accounts.md in the repository.
Task strategies. A chat model's catalog entry can now name the strategy it does its best task work under (tuning.taskStrategy), and the auto task strategy uses it for the model that will run the task. An explicit generalist-mode setting still wins. See docs/task-strategies.md in the repository.
Chat events. The usage chat event adds ttftMs, promptTokensPerSec and cachedPromptTokens where the engine reports them, including through a paired server.
Native engines are pinned to release 0.1.57, with a newer llama.cpp and the gezel-sandbox helper that confines no-network scripts in a container.
Content. The catalog is @bendyline/gilde 0.3.5, and the document editor is Squisq 2.13.10.

Things to see and do after update

Open Settings → Image generation and look at Styles and fast mode. If the model you draw with has a fast mode, install it, and pick a style from the strip the next time you ask for a picture. To change a picture you already have, ask the Meester for the Image editor. And if you want to know how well your own computer does with the model you use, open Settings → Benchmarks or run gezel bench. If you run a small business, open New Project and look through the new project types. To run gezel on a server instead, start with Gezel in Docker.