wanna.sh 0.8.1

• Agents' programs that exit cleanly have finished: they're no longer restarted over and over as if they had crashed.
  A program that crashes wakes its agent the first time it's restarted, not every time.
• Agents are told that their programs can't use their tools (such as web search), and to use a schedule for work that
  needs them, instead of writing programs that can't work.
• Agents' schedules repeat at most every 5 minutes, so one can't wake its model every minute all day.
• The Agents page warns when an agent's model is small (under about 4B): small models struggle to work on their own.
  Chat models' sizes come from their files.
• An agent's error no longer reads "The model didn't answer: ." when the problem came without a message.

Get it

Windows: download wanna.sh-setup-windows-x64.exe below and open it. It isn't code-signed yet, so if
Windows says "Windows protected your PC", click More info, then Run anyway.

Mac (Apple Silicon, macOS 14 or newer): download wanna.sh-macos-arm64.dmg below, open it and drag
wanna.sh into Applications. It isn't signed by Apple yet, so the first time, macOS says it couldn't verify
it: open System Settings → Privacy & Security, click Open Anyway, then Open.

Linux (or an Intel Mac):

curl -fsSL https://raw.githubusercontent.com/wanna-sh/wanna/main/install.sh | WANNA_BRANCH=v0.8.1 bash

Or with Docker: ghcr.io/wanna-sh/wanna:0.8.1 (everything) or ghcr.io/wanna-sh/wanna:0.8.1-chat (chat and Code only).
See the install guide and Docker guide.

wanna.sh 0.8.0

• Agents: an AI that runs around the clock. The new Agents page runs AIs that keep working on their own toward
  goals you give them, on your own chat model. Each has a Linux sandbox of its own (the same as Code's), where it
  writes and runs programs and keeps them running as services: wanna.sh restarts them when they stop and after
  restarts. It waits until something happens, or keeps thinking if you choose Always thinking, and wakes for your
  messages, schedules it sets itself, files its own programs write to /workspace/.inbox, and a webhook link other
  apps can post to. It tells you things with notifications: on its page, as a badge in the menu, on this computer, and
  on your phone through ntfy, or in Discord or Slack. It summarises its memory when it fills and keeps notes in
  MEMORY.md. Secrets are given to its services as environment variables and hidden from the model. A limit on steps
  an hour, help when it repeats itself and retries after failures keep it on track, and agents that were on come back
  on when wanna.sh starts. /api/agents does all of this from apps.

Get it

Windows: download wanna.sh-setup-windows-x64.exe below and open it. It isn't code-signed yet, so if
Windows says "Windows protected your PC", click More info, then Run anyway.

Mac (Apple Silicon, macOS 14 or newer): download wanna.sh-macos-arm64.dmg below, open it and drag
wanna.sh into Applications. It isn't signed by Apple yet, so the first time, macOS says it couldn't verify
it: open System Settings → Privacy & Security, click Open Anyway, then Open.

Linux (or an Intel Mac):

curl -fsSL https://raw.githubusercontent.com/wanna-sh/wanna/main/install.sh | WANNA_BRANCH=v0.8.0 bash

Or with Docker: ghcr.io/wanna-sh/wanna:0.8.0 (everything) or ghcr.io/wanna-sh/wanna:0.8.0-chat (chat and Code only).
See the install guide and Docker guide.

wanna.sh 0.7.0

• Train your own picture style. The new Train page teaches a Stable Diffusion 1.5 or SDXL picture model a
  subject (a person, a pet, a thing) or a style from 3 to 50 of your pictures, on your own graphics card. Choose
  what to teach, drop in pictures, name it, and press Start training: it suggests a trigger word and the number of
  steps, estimates how long it will take, and shows sample pictures when it's done. Use it in Pictures opens
  Pictures with it chosen; afterwards pick it under Style, with a strength setting under More settings.
• Trained styles in the API. /v1/images/generations and /api/generate/image take lora (a trained style's
  id) and lora_strength, and /api/training starts, follows and manages training from apps.
• Train a chat model. The Train page's Chat side teaches a chat model your writing voice (from your writing:
  text, Markdown or PDF files, or pasted) or how to answer (from example chats: typed in, or .jsonl, .json or .csv
  files). It trains on the model's original version, downloaded once, at full precision when it fits on your
  graphics card and in 4 bits (QLoRA) when it doesn't, across several cards if needed. The result is a chat model of
  its own in Chat, Code and the API, running in llama.cpp on a copy of the model made from the same original, with
  sample replies before and after on the Train page. /api/training/chat starts it from apps.
• Training needs more libraries, PEFT and bitsandbytes. The Train page offers to install them if your AI libraries
  came before this version.
• Your own AI for a task. The new Tasks page teaches the AI a job of your own from examples: flag, sort or
  score things from a table (gradient-boosted trees, in seconds, no graphics card needed), sort text or pictures
  (with your search model), or write answers your way and pull details out as JSON (a trained chat model). wanna.sh
  picks the method, tests it on examples it didn't learn from, and says how well it does in plain words, next to what
  always giving the most common answer would get; for yes/no tasks, a table of how many it catches against false
  alarms sets how sure it must be. A chat model can draft missing answers for you to check. Use it on the page, on a
  whole file, from /api/tasks, or as a tool the AI uses in Chat and Code.
• The AI libraries install now skips libraries meant for other systems.

Tested on Windows with an RTX 3060 12 GB: Stable Diffusion 1.5 (400 steps in under 4 minutes, 3.4 GB of graphics
memory) and SDXL (400 steps in about 13 minutes, 9.8 GB at most); Llama 3.2 1B (writing voice, under 2 minutes with
its download, 2.8 GB) and Qwen3 8B in 4 bits (30 example chats, about 4 minutes, 7.0 GB). Tasks: a standard breast
tumour table (96.7% right, against 62.7% for always giving the most common answer), 600 newsgroup posts in 4 topics
(94.3%, with Qwen3 VL Embedding 2B), cat, dog and car pictures, a made-up card-fraud table, and written answers with
Llama 3.2 1B; drafting with Qwen3 8B and Gemma 4 E4B. Not yet tested on Linux. Training isn't available on Macs yet.

Get it

Windows: download wanna.sh-setup-windows-x64.exe below and open it. It isn't code-signed yet, so if
Windows says "Windows protected your PC", click More info, then Run anyway.

Mac (Apple Silicon, macOS 14 or newer): download wanna.sh-macos-arm64.dmg below, open it and drag
wanna.sh into Applications. It isn't signed by Apple yet, so the first time, macOS says it couldn't verify
it: open System Settings → Privacy & Security, click Open Anyway, then Open.

Linux (or an Intel Mac):

curl -fsSL https://raw.githubusercontent.com/wanna-sh/wanna/main/install.sh | WANNA_BRANCH=v0.7.0 bash

Or with Docker: ghcr.io/wanna-sh/wanna:0.7.0 (everything) or ghcr.io/wanna-sh/wanna:0.7.0-chat (chat and Code only).
See the install guide and Docker guide.

wanna.sh 0.6.0

• Search pictures. With a search model that can see pictures, such as the new suggestion Qwen3 VL Embedding 2B,
  Knowledge libraries take photos and screenshots, and PDFs of scanned pages (each page is read as a picture).
  Describe what you're looking for to find matching pictures, or search with a picture to find ones like it. Chat
  shows the pictures it found, and gives them to a chat model that can see.
• Pictures in the embeddings API. /v1/embeddings takes pictures as image_url parts (data URLs) next to text,
  for models that can see them, and /api/knowledge/search takes an image to search with.

Tested on Windows with Qwen3 VL Embedding 2B and Gemma 4 E4B, with a small library of photos, a scanned PDF and a
note. Not yet tested on Macs or Linux, or with large picture libraries.

Get it

Windows: download wanna.sh-setup-windows-x64.exe below and open it. It isn't code-signed yet, so if
Windows says "Windows protected your PC", click More info, then Run anyway.

Mac (Apple Silicon, macOS 14 or newer): download wanna.sh-macos-arm64.dmg below, open it and drag
wanna.sh into Applications. It isn't signed by Apple yet, so the first time, macOS says it couldn't verify
it: open System Settings → Privacy & Security, click Open Anyway, then Open.

Linux (or an Intel Mac):

curl -fsSL https://raw.githubusercontent.com/wanna-sh/wanna/main/install.sh | WANNA_BRANCH=v0.6.0 bash

Or with Docker: ghcr.io/wanna-sh/wanna:0.6.0 (everything) or ghcr.io/wanna-sh/wanna:0.6.0-chat (chat and Code only).
See the install guide and Docker guide.

wanna.sh 0.5.0

• Your own tools. Write a Python function on the new Tools page (or as a file in ~/.wanna/tools) and the AI
  can use it: in Chat with the new Tools switch, in Code alongside the agent's own tools, and from apps by adding
  "wanna_tools": true to a chat request. The function's docstring tells the model what it does and its type hints
  what to give it; packages come from the standard # /// script block. Test a tool with your own inputs before the
  AI uses it.
• Tools run in a Docker sandbox (the internet, but not your files, your local network or other tools). One you trust
  can run directly on this computer instead; editing it takes the trust away, and in Code the agent asks before using
  it.

Tested on Windows with Docker Desktop, in Chat, Code and the API, with Gemma 4 E4B. Not yet tested on Macs or Linux.

Get it

Windows: download wanna.sh-setup-windows-x64.exe below and open it. It isn't code-signed yet, so if
Windows says "Windows protected your PC", click More info, then Run anyway.

Mac (Apple Silicon, macOS 14 or newer): download wanna.sh-macos-arm64.dmg below, open it and drag
wanna.sh into Applications. It isn't signed by Apple yet, so the first time, macOS says it couldn't verify
it: open System Settings → Privacy & Security, click Open Anyway, then Open.

Linux (or an Intel Mac):

curl -fsSL https://raw.githubusercontent.com/wanna-sh/wanna/main/install.sh | WANNA_BRANCH=v0.5.0 bash

Or with Docker: ghcr.io/wanna-sh/wanna:0.5.0 (everything) or ghcr.io/wanna-sh/wanna:0.5.0-chat (chat and Code only).
See the install guide and Docker guide.

wanna.sh 0.4.0

• Chat models bigger than memory. Turn on Settings → Run models bigger than memory and chat models bigger than
  your graphics and system memory together run by reading from your drive as they write, the way llama.cpp maps a
  model file. It's very slow, and works best with mixture-of-experts models such as Kimi K3 or DeepSeek, which read
  only a few of their experts for each token. Get models rates these Runs from your drive, very slowly whether
  or not the setting is on (saying how to turn it on), and before you download one you have to confirm you understand.
  Models with large tables llama.cpp reads a little at a time (such as Qwen3.8 Flash Next) are judged by what they
  really keep in memory, so they aren't mistaken for models that need the drive.
• Speed estimates before you download. For chat models, the version you pick shows how fast it will write on this
  computer, worked out from the model's own file (read from Hugging Face without downloading it) and measured speeds
  of your graphics cards, memory and drive. For a model that reads from the drive, it also says how long reading your
  message and starting take. My models shows the same under Options.
• Compression is explained. Chat versions show their bits per weight, and picking a heavily compressed one (below
  3 bits) says what that costs. Versions such as TQ1_0 and Q1_0 are now recognised as Tiny.
• Downloads that don't fit in the free space on your drive are refused up front, saying how much room they need.

Running from the drive was tested on Windows with NVIDIA graphics cards and under WSL2 (from WSL's own disk and from
a Windows drive), with Qwen3-30B-A3B at full precision (61 GB on a PC with 32 GB of memory). It hasn't been tested on
Macs or on Linux outside WSL yet, where estimates may be off. Estimates for models as big as Kimi K3 are worked out,
not measured.

Fixes

• The memory guard could stop a model that was only reading its file from the drive: on Linux it counted the file's
  pages as the model's own memory, and on Windows it counted memory the graphics driver had reserved but never used.
• A mixture-of-experts chat model too big for one graphics card could be put on the smallest card alone, with most
  of its experts left on the processor, when sharing it across cards was faster. It's now shared: Qwen3-30B-A3B at
  full precision went from 3 to about 4.7 tokens a second on an 8 GB and a 12 GB card.
• The speed My models measured for a chat model could be how fast it read a long message rather than how fast it
  wrote, which for models that run partly on the processor is often much higher.
• Mixture-of-experts chat models too big for the graphics cards could run out of graphics memory: wanna.sh passed
  llama.cpp's planned layer count without the plan's choice of which experts stay with the processor. Both are passed
  now.

Get it

Windows: download wanna.sh-setup-windows-x64.exe below and open it. It isn't code-signed yet, so if
Windows says "Windows protected your PC", click More info, then Run anyway.

Mac (Apple Silicon, macOS 14 or newer): download wanna.sh-macos-arm64.dmg below, open it and drag
wanna.sh into Applications. It isn't signed by Apple yet, so the first time, macOS says it couldn't verify
it: open System Settings → Privacy & Security, click Open Anyway, then Open.

Linux (or an Intel Mac):

curl -fsSL https://raw.githubusercontent.com/wanna-sh/wanna/main/install.sh | WANNA_BRANCH=v0.4.0 bash

Or with Docker: ghcr.io/wanna-sh/wanna:0.4.0 (everything) or ghcr.io/wanna-sh/wanna:0.4.0-chat (chat and Code only).
See the install guide and Docker guide.

wanna.sh 0.3.4

• wanna.sh is now copyright Wannaverse. The app's footer links to wannaverse.com, and it's
  still open source under the MIT licence.

Get it

Windows: download wanna.sh-setup-windows-x64.exe below and open it. It isn't code-signed yet, so if
Windows says "Windows protected your PC", click More info, then Run anyway.

Mac (Apple Silicon, macOS 14 or newer): download wanna.sh-macos-arm64.dmg below, open it and drag
wanna.sh into Applications. It isn't signed by Apple yet, so the first time, macOS says it couldn't verify
it: open System Settings → Privacy & Security, click Open Anyway, then Open.

Linux (or an Intel Mac):

curl -fsSL https://raw.githubusercontent.com/wanna-sh/wanna/main/install.sh | WANNA_BRANCH=v0.3.4 bash

Or with Docker: ghcr.io/wanna-sh/wanna:0.3.4 (everything) or ghcr.io/wanna-sh/wanna:0.3.4-chat (chat and Code only).
See the install guide and Docker guide.

wanna.sh 0.3.3

• Chat can search the web. With Web on, models that can use tools search the web and read pages when a
  question needs current or specific information, show what they're doing, and list the pages they read. It's on by
  default and can be turned off per chat.

Get it

Windows: download wanna.sh-setup-windows-x64.exe below and open it. It isn't code-signed yet, so if
Windows says "Windows protected your PC", click More info, then Run anyway.

Mac (Apple Silicon, macOS 14 or newer): download wanna.sh-macos-arm64.dmg below, open it and drag
wanna.sh into Applications. It isn't signed by Apple yet, so the first time, macOS says it couldn't verify
it: open System Settings → Privacy & Security, click Open Anyway, then Open.

Linux (or an Intel Mac):

curl -fsSL https://raw.githubusercontent.com/wanna-sh/wanna/main/install.sh | WANNA_BRANCH=v0.3.3 bash

Or with Docker: ghcr.io/wanna-sh/wanna:0.3.3 (everything) or ghcr.io/wanna-sh/wanna:0.3.3-chat (chat and Code only).
See the install guide and Docker guide.

wanna.sh 0.3.2

• The desktop apps update themselves. When a new version is out, the Windows and Mac apps offer to install it
  when they start, then reopen. Updates are signed and checked before they're installed. Apps from 0.3.1 or earlier
  need one last update by hand.

Get it

Windows: download wanna.sh-setup-windows-x64.exe below and open it. It isn't code-signed yet, so if
Windows says "Windows protected your PC", click More info, then Run anyway.

Mac (Apple Silicon, macOS 14 or newer): download wanna.sh-macos-arm64.dmg below, open it and drag
wanna.sh into Applications. It isn't signed by Apple yet, so the first time, macOS says it couldn't verify
it: open System Settings → Privacy & Security, click Open Anyway, then Open.

Linux (or an Intel Mac):

curl -fsSL https://raw.githubusercontent.com/wanna-sh/wanna/main/install.sh | WANNA_BRANCH=v0.3.2 bash

Or with Docker: ghcr.io/wanna-sh/wanna:0.3.2 (everything) or ghcr.io/wanna-sh/wanna:0.3.2-chat (chat and Code only).
See the install guide and Docker guide.

wanna.sh 0.3.1

Improvements

• Downloads carry on in the background. While a model downloads, wanna.sh keeps the computer from going to sleep
  (the screen still locks and turns off as usual), and if the connection drops, for example when the computer wakes
  up, it tries again by itself instead of failing.
• Saving says so. Save buttons show "Saved" for a moment, so one click makes one copy.
• Logs you can share. When a model fails, its log has Copy and Save buttons.
• Get models shows each rating as a small gauge, and drops the "Tested" labels.

Fixes

• Picture models downloaded in half precision (most of them, such as Stable Diffusion inpainting) failed to start on
  the processor with "no file named diffusion_pytorch_model.bin". They now load and run in full precision there.
• Big downloads on slower connections could fail with "The download kept stalling" although they were working:
  Hugging Face's downloader writes files in bursts, and long gaps looked like a stall. wanna.sh now watches the bytes
  arriving instead.
• On computers with swap, models were stopped as "ran out of system memory" while swap was still free. Free swap now
  counts, so they finish, more slowly.

Get it

Windows: download wanna.sh-setup-windows-x64.exe below and open it. It isn't code-signed yet, so if
Windows says "Windows protected your PC", click More info, then Run anyway.

Mac (Apple Silicon, macOS 14 or newer): download wanna.sh-macos-arm64.dmg below, open it and drag
wanna.sh into Applications. It isn't signed by Apple yet, so the first time, macOS says it couldn't verify
it: open System Settings → Privacy & Security, click Open Anyway, then Open.

Linux (or an Intel Mac):

curl -fsSL https://raw.githubusercontent.com/wanna-sh/wanna/main/install.sh | WANNA_BRANCH=v0.3.1 bash

Or with Docker: ghcr.io/wanna-sh/wanna:0.3.1 (everything) or ghcr.io/wanna-sh/wanna:0.3.1-chat (chat and Code only).
See the install guide and Docker guide.