v1.0.1

• docs: pull :latest again, now that a release publishes it
• ci: build the image for a release CI actually tagged
• server: stop the cooldown test racing the shed it measures
• docs: point the pull command at a tag that exists
• docker: name the actual reason a GPU is absent on Windows
• docker: say what a GPU does and does not reach in a container
• docker: one image that serves, backends downloaded not compiled
• docker: drop ik_llama.cpp from the unified image
• docker: make the unified image build fit inside CI

Full changelog: v1.0.0...v1.0.1

Instructions

Download the setup program for your platform:

• Windows -- quartermaster-setup-windows-amd64-v1.0.1.exe
• Linux -- quartermaster-setup-linux-amd64-v1.0.1 (or -linux-arm64-v1.0.1)
• macOS, Apple silicon -- quartermaster-setup-darwin-arm64-v1.0.1

The unix ones need chmod +x first. There is no signed macOS build yet, so
clear the quarantine flag before the first launch:

xattr -d com.apple.quarantine ./quartermaster-setup-darwin-arm64-v1.0.1

SHA256SUMS covers every file here.

The other files, and who they are for

quartermaster-windows-amd64.exe and quartermaster-linux-amd64 /
-linux-arm64 / -darwin-arm64 are the bare server binary, not an
installer. Nothing about them creates a Start Menu entry, a desktop shortcut or
an uninstall record. They are published for two reasons:

• the in-app updater downloads one to replace the running binary, so a release
  without them silently stops every existing install from updating;
• a headless Linux or macOS box can run one directly: chmod +x, point it at
  a models folder, and install backends from Settings.

Installing for the first time? Take a quartermaster-setup-... file from
the instructions above.

v1.0.0

Quartermaster's first public release.

Run any model without tuning a single flag. Point it at your models folder: every GGUF is
identified from its own header, the VRAM that is actually free is measured, and context length,
GPU offload, CPU-MoE split and KV sizing are computed per model. Models then hot-swap on demand
behind one OpenAI- and Anthropic-compatible API.

• Text, image and audio in one catalog, served by the inference servers you already trust:
  llama-server, stable-diffusion.cpp, whisper.cpp, vLLM, TabbyAPI, anything that speaks HTTP.
• One scheduler behind every listener, so everything loaded evicts against a single shared
  VRAM budget instead of colliding.
• A dashboard and a playground: live VRAM and throughput, per-model tuning, plus chat,
  images, speech and transcription.
• A Hugging Face browser and managed backends, so a model or a new llama.cpp build is a
  click rather than an afternoon.
• Self-updating from here on: later releases install themselves from inside the app.

Runs on your hardware. Bring your own models. No telemetry.

Instructions

Download the setup program for your platform:

• Windows -- quartermaster-setup-windows-amd64-v1.0.0.exe
• Linux -- quartermaster-setup-linux-amd64-v1.0.0 (or -linux-arm64-v1.0.0)
• macOS, Apple silicon -- quartermaster-setup-darwin-arm64-v1.0.0

The unix ones need chmod +x first. There is no signed macOS build yet, so
clear the quarantine flag before the first launch:

xattr -d com.apple.quarantine ./quartermaster-setup-darwin-arm64-v1.0.0

SHA256SUMS covers every file here.

The other files, and who they are for

quartermaster-windows-amd64.exe and quartermaster-linux-amd64 /
-linux-arm64 / -darwin-arm64 are the bare server binary, not an
installer. Nothing about them creates a Start Menu entry, a desktop shortcut or
an uninstall record. They are published for two reasons:

• the in-app updater downloads one to replace the running binary, so a release
  without them silently stops every existing install from updating;
• a headless Linux or macOS box can run one directly: chmod +x, point it at
  a models folder, and install backends from Settings.

Installing for the first time? Take a quartermaster-setup-... file from
the instructions above.

Runtime assets v1

Static assets quartermaster downloads at runtime rather than shipping in the binary. Not a software release — do not delete, the URLs here are pinned in source.

titlegen-flan-t5-small-q8_0.gguf (79 MiB) — the chat-title model. Pinned by URL and SHA-256 in internal/server/titlegen_asset.go; fetched on first use and prefetched by the Windows installer. Chat titles fall back to the chat model when it is unavailable.

sha256  a040e12a77a3da86491a4347296cfd16b76b41e6c6b19b58fe6f2dc072edccb9
size    82866336

Derived from google/flan-t5-small (Apache-2.0) via llama.cpp convert_hf_to_gguf.py (f16) then llama-quantize to Q8_0.