Ollama v0.32.7 リリースOllama Releases v0.32.7

Ollama has published version 0.32.7, a point release whose headline change is initial support for Muse Glimmer, described in the release notes as Meta's newest open model. For people who run large language models on their own hardware, the update is notable because it makes a fresh Meta model available through Ollama's tooling almost immediately, though for now that access is limited to a specific hardware path.

According to the release, Muse Glimmer is currently available via initial support through Ollama's MLX engine on Apple Silicon. In other words, the first way to run the model inside Ollama is on Apple's own chips, using the MLX-based execution backend rather than the project's more traditional inference paths. The release also states that additional support and optimizations for Apple Silicon, NVIDIA, AMD, and other platforms will be available in the coming days, signaling that the Apple Silicon launch is a starting point rather than the full rollout.

The reliance on MLX is the key technical detail here. MLX is Apple's open-source array and machine learning framework built specifically for Apple Silicon, designed to take advantage of the unified memory architecture shared between the CPU and GPU on M-series chips. Ollama has been expanding beyond its longstanding llama.cpp foundation to incorporate an MLX engine, which allows certain models to run with Apple-native acceleration. Launching Muse Glimmer through this engine first suggests the model has been prepared or converted in a form that the MLX path can handle efficiently, while the broader engine work needed for NVIDIA and AMD GPUs appears to still be in progress.

The staggered platform availability is worth reading carefully rather than assuming day-one parity everywhere. Users on NVIDIA or AMD systems, which typically depend on CUDA or ROCm acceleration, will need to wait for the promised follow-up support before they can run Muse Glimmer locally through Ollama. The phrase "in the coming days" indicates the gap is intended to be short, but the release does not give a firm date, so timing is likely to depend on further updates.

For background, Ollama is a widely used tool for running open models locally, packaging model downloads, quantized weights, and a simple command-line and API interface into a single application that works across macOS, Linux, and Windows. It has become a common on-ramp for developers and enthusiasts who want to run models without relying on cloud services, and its model library often adds new open releases shortly after they appear. The project competes with and complements other local-inference tools such as LM Studio, Jan, and the underlying llama.cpp runtime that many of these front ends share.

Meta's continued release of open models forms the other half of the story. The company's Llama family has been central to the open-weight ecosystem, and each new Meta model tends to be quickly adopted by runtimes like Ollama because of broad community demand and tooling compatibility. Muse Glimmer is presented here as Meta's newest open model, and its arrival in Ollama so soon after release fits the established pattern of rapid integration. The source excerpt does not detail the model's size, architecture, licensing terms, or benchmark performance, so those specifics remain to be confirmed from Meta's own documentation.

The choice to lead with an MLX-on-Apple-Silicon launch also reflects a broader industry trend toward optimizing local inference for consumer hardware. Apple Silicon's unified memory can make it practical to load sizable models on laptops and desktops that would otherwise require discrete GPUs with large amounts of dedicated VRAM, and MLX has become an increasingly important framework for squeezing performance out of those chips. Prioritizing that path can get a model into users' hands quickly while the more complex, multi-vendor GPU support is finalized.

In practical terms, anyone wanting to try Muse Glimmer today will need an Apple Silicon Mac and an up-to-date Ollama installation running version 0.32.7. Those on other platforms may prefer to wait for the announced NVIDIA, AMD, and additional Apple Silicon optimizations before pulling the model. As with any early or initial support, some rough edges or performance limitations are possible until the wider platform work and optimizations described in the release notes land.