Ollama released v0.32.7 on August 10, 2026, and the headline is a model rather than a feature. Muse Glimmer is a new open model from Meta, the first one published by Meta Superintelligence Labs, and it is now available through Ollama.
Read the caveat before you clear disk space for it, because it matters more than the announcement does for most of us. Support in 0.32.7 is initial, and it only works on Ollama’s MLX engine on Apple Silicon. If your AI box is an x86 machine with an NVIDIA card, which describes most homelabs, this release gives you nothing to run yet.
What Muse Glimmer is
Muse Glimmer is a 30B multimodal model that Meta built specifically for agent workloads running locally. That last part is what makes it interesting for self-hosting. Plenty of open models are good at chat and mediocre the moment you ask them to call tools in a loop. A model designed around local agent work is a different proposition, at least on paper.
Multimodal here means image input, which Ollama’s MLX engine supports starting with 0.32.7. The engine also handles DFlash for this model.
The Apple Silicon limitation, spelled out
Ollama’s own release notes describe the current support as initial. Additional support and optimizations for Apple Silicon, NVIDIA, AMD and other platforms are described as arriving in the coming days. No firm date, no specific build number.
So the practical situation on August 14 is this. A Mac with an M-series chip and enough unified memory can run Muse Glimmer today. Everything else waits. If you have been planning a GPU purchase around this model, wait for the platform support to actually land rather than buying on a roadmap sentence.
Ollama says its MLX engine delivers state-of-the-art performance on Apple Silicon for this model. That claim is theirs, not a benchmark I have reproduced, and it is worth treating as a vendor statement until independent numbers show up.
Pulling and running the model
The documented command is short:
ollama run muse-glimmer:30b-mlxThe -mlx suffix in the tag is not decoration. It tells you exactly which engine this build targets, and it is a useful reminder of why the tag will not help you on a Linux server with a 3090 in it.
Wiring it into coding agents and assistants
The part I find more interesting than the model itself is the set of ollama launch commands shipped alongside it. Ollama documents Muse Glimmer as a backend for several agent applications, all running entirely on your own hardware:
ollama launch claude --model muse-glimmer:30b-mlxfor Claude Codeollama launch pi --model muse-glimmer:30b-mlxfor Pi, a lighter coding agent from pi.devollama launch openclaw --model muse-glimmer:30b-mlxfor OpenClawollama launch hermes --model muse-glimmer:30b-mlxfor Hermes
OpenClaw and Hermes are positioned as long-running personal assistants rather than coding tools. Ollama also lists Codex among the coding agent applications it can serve. The common thread is that the model stays on your machine, so the code your agent reads and the context it accumulates never leave the network.
This continues a direction we covered last month, when Ollama turned its bare CLI into an interactive agent by default. The pattern is consistent: Ollama is spending its releases on agent plumbing rather than on the chat experience.
Should you act on this
If you run a Mac Studio or a well-specced MacBook as your local inference box, this is a reason to update Ollama this week and see how a 30B multimodal model behaves against your actual agent workflows. Test it on a real task, not a demo prompt.
Everyone else can update Ollama for the usual reasons and keep an eye on the release feed. There is no downside to being on 0.32.7, but there is no Muse Glimmer for you in it either.
If you are still setting up the underlying stack, our Ollama application page covers the basics, and Open WebUI remains the simplest way to put a browser interface in front of whichever model you end up running.
Full details are in the official v0.32.7 release notes on GitHub.