FreeToken Desktop v0.2.0-beta.22

What's Changed

Features

• freetoken command on Windows, for the terminal: in PowerShell ft is the built-in Format-Table alias. Available after the next engine install or update.

Fixes

• A model that runs out of system memory while loading is now reported as out of memory, with advice to close apps or enlarge the Windows page file, instead of as damaged weights. Needs the latest engine (update it in Settings).
• The engine install scripts pin the uv version.
• Linux: install.sh pins the flashinfer version, fixing FlashML-org/FreeToken#543.

FreeToken Desktop v0.2.0-beta.21

What's Changed

Features

• User-friendly error when a model's weights fail to load, with a guide to open a GitHub issue. Needs the latest engine (update it in Settings).

Fixes

• Linux: the engine install no longer fails resolving sglang-kernel.

FreeToken Desktop v0.2.0-beta.20

What's Changed

Fixes

• Fixed the FTW repair check: some checkpoints were flagged for a repair they did not need.
• Fixed FTW repair compatibility with proxies

FreeToken Desktop v0.2.0-beta.19

What's Changed

Features

• Chat attachments: drop, paste or pick images, text files and PDFs straight into a conversation.
• VLM support: vision models now take images and answer on them. Needs engine 0.1.3 or newer — see which families it covers.
• Some FTW checkpoints converted by an earlier FreeToken need a repair before the updated engine loads them — the Library does it in place, without downloading the model again.

Fixes

• On Linux, a self-update through deb or pacman no longer hangs on a password prompt you could not see.
• API request times in Logs now read in your own time zone.
• A long engine error no longer stretches the failure dialog past the edge of the window; it scrolls instead.

FreeToken Desktop v0.2.0-beta.18

What's Changed

Features

• Proxy settings: System, Off, or a custom http/https/socks5 proxy with optional credentials.
• UI: Settings now has a side menu to jump to any section.

Fixes

• Engine install and update checks now trust system certificates.
• On Linux, the settings file holding your Hugging Face token is now readable only by your user.

FreeToken Desktop v0.2.0-beta.17

What's Changed

Features

• Settings page links to the changelog.
• Console restores each model's last applied cache config on load.

Fixes

• Engine install now resolves the wheel pair from a per-platform manifest on the FreeToken-Web beta release, falling back to the GitHub API only when the manifest is missing. Fixes the 403 rate-limit failures on unauthenticated installs.
• Switching models waits for the stopped serve's port to be released instead of showing an error banner.
• Daemon port moved to 19000 with a bind test at startup; a legacy daemon left running is adopted instead of colliding.
• Library shows an error instead of a stand-in VRAM figure when the local GPU cannot be read.
• NVML now loads via libnvidia-ml.so.1, fixing GPU detection on Fedora and other distros without the unversioned symlink.

FreeToken Desktop v0.2.0-beta.16

What's Changed

Features

• Qwen3.8-Flash-Next NVFP4 and GLM-5.3-Flash NVFP4 added to the model catalog
• Downloads no longer auto-convert to FTW. Start a model directly; convert when you want faster loads

Fixes

• Download progress based on real size, not hardcoded
• Windows: the engine installer no longer edits your PATH
• Linux: the engine installer no longer picks up an unrelated uv on PATH
• Linux/Wayland: fixed a blank window on recent Mesa and a startup crash on NVIDIA

FreeToken Desktop v0.2.0-beta.15

What's Changed

Features

• Multi-GPU select support — Settings → GPU lists every NVIDIA card and lets you choose which one the engine runs on. (Multi-GPU inference is still on the roadmap and not yet available.)
• Hardware check on launch — shows a message if there is no NVIDIA driver or the driver is too old

Fixes

• A FTW conversion that has died or stalled no longer looks like a working one
• Fixed the fit estimate for dense models
• AppImage: helper processes (installer, engine, converter) no longer inherit the AppImage's library paths
• Windows engine install no longer depends on a uv already on the box
• Gemma-4-26B-A4B BF16 card now downloads the instruct checkpoint, https://huggingface.co/google/gemma-4-26B-A4B-it, instead of the pretrained base
• Some UI fixes

FreeToken Desktop v0.2.0-beta.14

What's Changed

Features

• A model's quantizations now stack into one Library card
• Qwen3.8-27B added to the model catalog (NVFP4, FP8, BF16)

FreeToken Desktop v0.2.0-beta.13

What's Changed

Features

• Launch DeepSeek Harness from the Apps screen.
• New icons and theme.
• Translated system tray menu.

Fixes

• A crashed engine blocked every later model start.
• The Apps install card hung after its terminal closed.