Skip to content

v1.6.0 - 2026-10-24

Choose a tag to compare

@github-actions github-actions released this 31 Oct 10:33
2a9961b

Release Notes

  • hedos no longer needs Ollama or uv installed first (#33). Everything a model needs to run is either carried in the binary or installed by hedos when you say yes. hedos never installs an engine on its own. docs/engines.md is a new guide to how hedos finds, installs and runs its engines.
  • llama.cpp comes with hedos. The install script, Homebrew and the release archives carry llama.cpp 0.7.0 (build 11669): the Metal build on Apple Silicon, the CPU build on Intel Macs and Linux. It unpacks into the data directory the first time a model runs, and hedos runs it ahead of any llama-server on your PATH, so an older llama.cpp there never gets in the way ([engines] llama_server picks another). Release archives grow from about 5 MB to between 12.8 and 15.5 MB. A cargo install or source build carries none, and hedos pull offers to install one once a GGUF model lands.
  • hedos engines lists the engines models run on (llama.cpp, Ollama and uv), where each was found, its version, and what to do when one is missing. hedos engines install llama.cpp fetches a faster build for the machine (CUDA for an NVIDIA card, Vulkan for AMD, or one you pick with --variant). hedos engines install uv fetches uv 0.13.4. Each install first says what it will fetch, from where, how big and where it goes, then asks, defaulting to no. Every archive is checked against its pinned size and SHA-256, and the program is run once before it is put in place. A Ctrl-C leaves nothing half-installed. hedos engines remove takes back what hedos installed, never one you installed yourself. When a faster build suits the machine, hedos engines and hedos recommend say so.
  • On Linux, hedos and the llama.cpp it carries run from glibc 2.35 (Ubuntu 22.04, Debian 12) with no other package. That build is hedos's own, built from llama.cpp's pinned source, and on an older system it is also the one hedos engines install fetches. llama.cpp's own GPU builds need glibc 2.38 (Ubuntu 24.04, Debian 13): on an older system, asking for one says why it would not start before anything is fetched.
  • Ollama is optional. The models in its store stay on the shelf: Ollama serves them while it is installed, and llama.cpp serves the ones it can load when it is not. Two kinds stay with Ollama: a model whose vision or audio encoder is packed into its weights, and one whose only chat template is Ollama's own. hedos ls and the shelf say those need Ollama. hedos rm deletes an Ollama model without Ollama too, keeping any blob another tag shares.
  • The install catalog is Hugging Face GGUF repos served by llama.cpp: Qwen3.5, Gemma 4, gpt-oss and Qwen3.8 for chat, Qwen2.5-Coder, Qwen3-Coder and Qwen3.6 for code. hedos recommend and both pull surfaces show them by their Hugging Face names. A GGUF pull no longer fetches the speculative-decoding drafts llama.cpp's repos ship beside a model. It keeps one projector, the one matching the weights' quantization.
  • A chat model's default window fits the memory the server can use as it starts: on Linux what is available inside hedos's cgroup, on a Mac what is free now. A 7B model now starts on an 8 GB machine instead of being killed while it loads. A server killed while loading is reported as most likely out of memory, with the window it started at.
  • uv comes from hedos when you have none. A Python model's first pull ends by offering hedos engines install uv. On the shelf, w or t on such a model opens the same install card. uv then fetches Python 3.12 into hedos's data directory when the machine has none. Every package is checked against the lock's hashes whatever your uv.toml says. A model whose environment is built keeps running if uv is removed.
  • The torch runtimes work on Linux: audio, diffusers images, laya, zerank and YuE2. Each failed there before. Each now ships a PyPI lock (the CUDA build and NVIDIA's libraries on Linux, Metal on a Mac) and a CPU-only lock beside it. hedos installs the CPU one on Linux when NVIDIA's driver is not loaded. A card the installed torch has no kernels for runs on the CPU instead of failing every job. The diffusers runtime uses full precision on the CPU. The CPU build needs uv 0.6.9 or later; an older uv is told to run hedos engines install uv. The first CUDA install fetches about 3.5 GB, and package downloads may now take up to two hours.
  • A manifest runtime's approval also covers its lockfile's CPU variant, wherever the manifest's path leads. Existing approvals still count, unless the manifest names its lockfile through .. or in another letter case than the file has, which asks for approval once more.
  • A split GGUF model from the Hugging Face cache loads: llama-server is started by its first shard's file name, which it needs to find the others.
  • A folder reached twice (a watched folder inside another, a symlink to one, or ~/models beside ~/Models on a disk that ignores case) gives its models once.
  • A prompt no longer ends with exit 130 by itself. When a child process exited while a question was open, such as the daily release check, the answer was lost as if Ctrl-C had been pressed. A first hedos pull of the day could lose its "Download now?" a second after showing it.

Install hedos 1.6.0

Install prebuilt binaries via Homebrew

brew install theiskaa/tap/hedos

Download hedos 1.6.0

File Platform Checksum
hedos-aarch64-apple-darwin.tar.xz Apple Silicon macOS checksum
hedos-x86_64-apple-darwin.tar.xz Intel macOS checksum
hedos-aarch64-unknown-linux-gnu.tar.xz ARM64 Linux checksum
hedos-x86_64-unknown-linux-gnu.tar.xz x64 Linux checksum