The fastest way to make old hardware feel new is to stop emulating things that never needed emulating.

When we first packaged Meta’s Muse Code for AliveOS, upstream’s x86_64 binary demanded AVX2. Anything pre-Haswell — Sandy Bridge, Ivy Bridge, early Atoms and Celerons, the machines AliveOS exists to keep alive — died on launch with Illegal instruction. So we did the pragmatic thing: detect the missing flag in /proc/cpuinfo and silently reroute through qemu-x86_64 -cpu max. It worked. It was also a tax on every single invocation, on machines that could least afford it.

This week that tax got repealed. Upstream killed the AVX2 baseline, and we rebuilt the packaging around that fact. Here is what changed, where it comes from, and why Muse Code is now our endorsed AI coding agent.

What changed

1. Upstream 1.4.10, verified native on pre-AVX2 silicon. Our repo carried 1.4.2; we jumped straight to 1.4.10-R6427.1. And we did not take “runs everywhere” on faith: our own build machine has AVX but no AVX2, and the new binary runs on it directly — --version, real exec workloads, exit zero, no emulator in the process tree. The original crash-at-startup failure mode is gone.

2. QEMU is now opt-in, not automatic. The wrapper executes natively by default. Emulation is still there for edge cases, but you have to ask for it explicitly:

MUSE_QEMU=1 muse
# or
muse --qemu

Why opt-in instead of auto-detect? Auto-detect was the right call when the binary could not run without emulation. Now it can, and automatic emulation would mean silently paying 2-5x startup overhead forever on exactly the hardware that should be spared. Defaults should cost nothing; exotic fallbacks should require a signature. If you ever do hit a real Illegal instruction on ancient silicon, install qemu-user and opt in. Otherwise forget QEMU exists.

3. AUR first, AliveOS second — one pipeline. muse-code-bin on the AUR is now the single source of truth: it gets the version bump and the wrapper changes first. The aliveos-repo build pulls from the AUR instead of tracking a separate upstream clone, and CI rebuilds and republishes the latest release from there. One packaging, reviewed once, shipped everywhere: AliveOS (sudo pacman -S muse-code — no -bin suffix needed, the repo package provides that name), AUR (yay -S muse-code-bin), Termux (aarch64 deb).

A correction to our own Android post: it told owners of x86_64 tablets without AVX2 to add qemu-user-x86-64. That advice is retired — current releases run natively there too. (proot on Android is still required; that is a seccomp problem, not a CPU problem, and it is staying.)

Upstream references and docs

Why we endorse it: cheap inference, zero-friction setup

AliveOS does not endorse tools lightly — endorsement here means “this is the default answer when someone asks what AI agent to run.” Muse Code earned it on two axes that matter more than benchmark scores:

Cheap inference. The model lives behind an API, not behind a GPU you have to buy, cool, and feed. For the old-hardware crowd this is the whole game: a Sandy Bridge box with 8 GB of RAM cannot run a serious local model, but it can stream diffs from Muse Spark all day. Metered inference turns AI assistance from a hardware purchase into a utility bill — and a small one. No VRAM, no quantized-model roulette, no watching your machine swap itself to death.

Easy setup. This is where our packaging earns its keep:

yay -S muse-code-bin      # Arch / AUR
sudo pacman -S muse-code  # AliveOS repo (no -bin suffix needed)
muse "fix the failing tests"   # ...that is the whole onboarding

Zero configuration for the cases that used to need forum threads: Termux proot wrapping on Android, TLS paths, working-directory handling. Plus two tools upstream does not ship — muse sessions, a full curses TUI for finding and resuming past work, and muse mcp import, which vacuums up MCP servers from Claude Code, Codex CLI, OpenCode, and Antigravity configs you already have. Switching agents should not mean rewiring your tools.

Terminal-first, no daemon, no Electron, no background wake-locks. It respects the battery, which on AliveOS is a moral issue, not a technical one.

Try it, on your worst machine

That is the actual test. Dig out the oldest x86_64 box you have, install from the repo, and run muse --version. If it prints a version instead of dumping core, the AVX2 era is officially behind us. If it doesn’t — file an issue with your /proc/cpuinfo flags and we will look at it personally. Old hardware is not an edge case here. It is the mission.