Releases: IterateAI/lifeboat-releases
Release list
Lifeboat Desktop v2.2.52
Lifeboat Desktop v2.2.52 — Windows, macOS (Apple Silicon) and Linux.
Downloads
| Platform | File |
|---|---|
| Windows 10/11 x64 | Lifeboat-2.2.52-setup.exe — signed installer, just run it |
| macOS Apple Silicon | Lifeboat-2.2.52-macos-arm64.dmg — notarized |
| macOS Intel | Lifeboat-2.2.52-macos-x86_64.dmg — notarized |
| Linux x86_64 | lifeboat-desktop_2.2.52_amd64.deb or Lifeboat-2.2.52-linux-x86_64.tar.gz |
| Linux arm64 | lifeboat-desktop_2.2.52_arm64.deb or Lifeboat-2.2.52-linux-aarch64.tar.gz |
Install the .deb with sudo apt install ./lifeboat-desktop_2.2.52_amd64.deb rather
than dpkg -i, so the graphics driver packages it recommends are pulled in.
What changed
Speculative decoding now works on the GGUF engine. The setting was offered on the
server form but had no effect there — the server started and simply never used it. On a
model that ships a built-in draft stage, enabling it measured 43% faster for a single
request and 72% faster with four at once, on the same hardware and the same model. Leave
Speculative Decoding on Auto and Lifeboat picks the best method the model file supports.
Fixed: Auto chose the weaker draft method for every GGUF model. It looked for the
model's built-in draft stage in a companion file that this format does not have, found
nothing, and fell back to the generic method — so speculation was on, but the slower
kind, with nothing in the log to say so. It now reads the model file itself. Existing
servers pick this up on their next start.
Fixed: models that reason before answering were reported as not supporting tool
calling. Lifeboat tests this by asking the model to call a tool, and the allowance was
small enough that such a model spent it all reasoning and never reached the call. Clients
reading that listing would stop offering tools to a model that supports them perfectly
well. The published capability corrects itself the first time each server starts on this
release.
Downloading a model into a folder Lifeboat cannot write to now fails immediately,
naming the folder, its owner and the command that fixes it. Previously the download
appeared to start, the model was marked as downloading, and it failed in the background
with a permissions code and a path.
Windows server starts. 2.2.51 fixed process handling that was written for macOS and
Linux only and made every server start fail on Windows; this release carries that fix and
is the first Windows build since to ship as a signed installer.
Maintenance and security release. The third-party components bundled in the product
have been updated, including several carrying security fixes — among them the library
that handles encryption and certificate validation. Nothing to configure and no action
required.
Verifying what you downloaded
The Windows installer and everything inside it are signed by Iterate Studio Inc
(DigiCert EV code-signing certificate, SHA-256, RFC 3161 timestamped), so Windows names
the publisher before it runs. The macOS disk image is signed, notarized by Apple and
stapled.
SHA-256 checksums for every file are in desktop/installers/RELEASES.json in the source
repository, and in SHA256SUMS-linux.txt attached here for the Linux artifacts.
Lifeboat Desktop 2.2.51 (Windows)
Lifeboat Desktop 2.2.51 — a Windows-only release.
On macOS or Linux? Download 2.2.50, which is current for those platforms. Only the Windows installer changed here, so there is no 2.2.51 build for them and nothing to upgrade.
Windows
| File | Notes |
|---|---|
Lifeboat-2.2.51-setup.exe |
Signed installer, no administrator prompt. Needs the free .NET 8 Desktop Runtime. |
Fixes: no inference server could be started, or stopped, on Windows.
Every POST /api/servers/{id}/start failed before the engine was even spawned, because the control plane called POSIX-only process APIs that do not exist on Windows:
os.waitpid(-1, os.WNOHANG)— the zombie reaper. Windows has no zombies and nowaitpid(-1); the missingos.WNOHANGraised before any handler saw it, and the same call ran on a 2-second loop filling the log.os.getpgid/os.killpg— the stop path, so a server could not have been stopped either.signal.SIGKILL— absent on Windows, whereSIGTERMis already the non-catchable kill.os.statvfs— the shared-memory preflight, fatal for any multi-GPU launch.
Also fixed: the per-server log path /tmp/lifeboat_server_<port>.log is drive-relative on Windows and resolved to a directory that need not exist. The four spawn sites now create it; the read side is unchanged so both still agree on the path.
Every change is a platform guard, so macOS and Linux behaviour is unchanged — which is why those platforms stay on 2.2.50 rather than being rebuilt.
Verified on Windows 11 against a real install: the control plane imports, a server starts, inference returns through the proxy, stop reaps the engine, and start succeeds again.
Checksums and provenance: desktop/installers/RELEASES.json.
Lifeboat Desktop 2.2.50
Windows users: get 2.2.51 instead. The 2.2.51 installer fixes a Windows-only defect that stopped any inference server from starting. macOS and Linux are current on this release.
Lifeboat Desktop 2.2.50 — official release for Windows, macOS (Apple Silicon and Intel) and Linux (x64 and arm64).
Run language models on your own machine: no container runtime, no root, and an OpenAI- and Anthropic-compatible API behind a local console.
Downloads
| Platform | File | Notes |
|---|---|---|
| Windows 10/11 x64 | Lifeboat-2.2.50-setup.exe |
Signed installer, no administrator prompt. Needs the free .NET 8 Desktop Runtime. |
| macOS Apple Silicon 13+ | Lifeboat-2.2.50-macos-arm64.dmg |
Metal + MLX. Signed, notarized and stapled. |
| macOS Intel 13+ | Lifeboat-2.2.50-macos-x86_64.dmg |
GGUF, no MLX. Signed, notarized and stapled. |
| Linux Debian/Ubuntu x64 | lifeboat-desktop_2.2.50_amd64.deb |
GPU offload via Vulkan. |
| Linux Debian/Ubuntu arm64 | lifeboat-desktop_2.2.50_arm64.deb |
GPU offload via Vulkan. |
| Linux any distro, x64 | Lifeboat-2.2.50-linux-x86_64.tar.gz |
GPU offload via Vulkan. |
| Linux any distro, arm64 | Lifeboat-2.2.50-linux-aarch64.tar.gz |
GPU offload via Vulkan. |
macOS and Windows builds are code-signed and checked by the operating system. Linux builds are not signed; verify them against SHA256SUMS-linux.txt:
sha256sum -c SHA256SUMS-linux.txt --ignore-missingFull instructions: docs/desktop.md
What is new
macOS Apple Silicon ships the right engine. 2.2.49 contained an arm64 app with an x86_64 inference engine. It still worked on a Mac that has Rosetta — measured about 13% slower to generate, with the GPU still doing the work — but on an Apple Silicon Mac without Rosetta it could not start at all. 2.2.50 is built natively for arm64 with Metal and the Apple CPU kernels.
macOS Intel is built again. Upstream stopped publishing Intel Mac packages for one of our dependencies, so this build compiles it from source against its own OpenSSL rather than shipping the last Intel-compatible release, which carries six published advisories.
Linux uses the GPU. Every previous Linux build ran every model on the CPU regardless of hardware, because the GPU backend was silently omitted. 2.2.50 offloads through Vulkan, covering NVIDIA, AMD and Intel with one download. Install the .deb with apt install ./file.deb rather than dpkg -i so the Vulkan loader is pulled in.
Windows ships a signed installer rather than a zip you unpack by hand, with a Start-menu entry, optional start-at-sign-in and a proper uninstaller.
Checksums and provenance for every artifact are in desktop/installers/RELEASES.json.
Lifeboat Desktop v2.2.49
Superseded by 2.2.50 — do not use this build on Apple Silicon.
The macOS arm64 disk image here contains an x86_64 inference engine. On a Mac with Rosetta it still runs (about 13% slower to generate); on an Apple Silicon Mac without Rosetta installed the engine cannot start at all. 2.2.50 is built natively for arm64.
Kept for reference and rollback only.
Signed desktop artifacts for v2.2.49. Files under 95 MB are committed to desktop/installers/; larger ones are attached here. See desktop/installers/RELEASES.json for the full list with checksums.
Lifeboat Desktop 2.2.46
Superseded by 2.2.50.
The Linux builds here have no GPU backend and run every model on the CPU regardless of hardware. Kept for reference and rollback only.
Lifeboat Desktop 2.2.46. Supersedes 2.2.45.
Three first-run defects, all reported from one Mac, all in the path a new
install takes on its first day.
Fixed — MLX models could not be downloaded at all
An MLX checkpoint is safetensors: mlx-community/Qwen3-8B-4bit holds
model.safetensors and no GGUF. The gate that refuses safetensors an install
cannot serve therefore matched it exactly and disabled Download, while its
own message advised looking for "an mlx-community build" — it was refusing the
thing it recommended.
The browser copy of that gate had been stricter than the server's, which
exempted MLX from the start. It now mirrors the server rather than carrying a
second rule.
Fixed — an MLX download was allowed where MLX cannot run
The name-based exemption was unconditional, so pasting an mlx-community
repository on an Intel Mac, on Windows, or in the Lite image began
a download that could never load. Both gates now require MLX to actually be
present. GGUF keeps no such condition — that engine ships in every shape.
Fixed — a recommended model pointed at a repository that does not exist
Qwen/Qwen3-4B-Instruct-2507-GGUF returns Repository Not Found — Qwen
publishes no GGUF for that model. Because the download dialog has a
HuggingFace token selector directly under the error, it reads as an
authentication problem; it was a wrong repository id in our catalogue. It now
points at a verified one.
Every repository Lifeboat recommends is existence-checked before release, and
checked against its own format tag — a repository that exists can still be the
wrong one.
Downloads
| Platform | File |
|---|---|
| macOS — Apple Silicon (M1–M4), 13+ | Lifeboat-2.2.46-macos-arm64.dmg |
| macOS — Intel, 13+ | Lifeboat-2.2.46-macos-x86_64.dmg |
| Linux — Debian/Ubuntu x64 | lifeboat-desktop_2.2.46_amd64.deb |
| Linux — Debian/Ubuntu arm64 | lifeboat-desktop_2.2.46_arm64.deb |
| Linux — any distro, x64 | Lifeboat-2.2.46-linux-x86_64.tar.gz |
| Linux — any distro, arm64 | Lifeboat-2.2.46-linux-aarch64.tar.gz |
Both Mac builds are signed with an Apple Developer ID, notarized by Apple and
stapled — they open with a double-click and work offline. The disk image
contains an Applications shortcut for drag-and-drop. Linux builds need
glibc 2.31 or newer (Debian 11+, Ubuntu 20.04+, RHEL 9+).
Verify with shasum -a 256 -c SHA256SUMS.
Container images for this release: https://hub.docker.com/r/iterateai/lifeboat
Lifeboat Desktop 2.2.45
Superseded by 2.2.46. Marked pre-release so releases/latest resolves to 2.2.46. On 2.2.45, MLX models could not be downloaded at all, and a recommended GGUF model pointed at a repository that does not exist — both fixed in 2.2.46.
Lifeboat Desktop 2.2.44
Superseded by 2.2.45. Marked pre-release so releases/latest resolves to 2.2.45. 2.2.44's Help links all resolve — they open the documentation entry page for each console page rather than the focused page, so nothing is broken; 2.2.45 closes that one click.
Lifeboat Desktop 2.2.43
Superseded by 2.2.44. Marked pre-release so releases/latest resolves to 2.2.44. In 2.2.43 an unservable model could still be selected for download and failed only after the click; 2.2.44 refuses it in the picker.
Lifeboat Desktop 2.2.42
Superseded by 2.2.43 — download that instead.
In 2.2.42, clicking Start on a server flickered between Starting and the
previous state, and each flicker closed an open log panel.
Lifeboat Desktop 2.2.41
Superseded by 2.2.42 — download that instead.
2.2.41 pinned a memory fraction on deployed servers (which could OOM), offered
models the machine cannot serve, and listed MLX above GGUF while saying GGUF
was the default.