Meta's 30B Muse Glimmer lands on Apple Silicon today via Ollama's MLX engine

A weight drop without a local runtime is homework; this one is a download. Meta opened the weights for Muse Glimmer, a 30B parameter dense model, and Ollama made it runnable the same day. That pairing is the story.

Ollama's post is direct: Muse Glimmer ships today through its MLX engine, with what the project calls state-of-the-art performance on Apple Silicon. It can power Claude Code, Codex, and other always-on local agent workflows natively. Mark Zuckerberg, announcing the release on X, described Glimmer as a "great 30B parameter dense model that can run locally" and said Meta would open the weights for Muse Spark 1.2, its latest foundation model, soon. "Meta is a strong supporter of open source and I'm proud of these releases," he wrote, echoing the open-source bet Meta has spent billions on.

The MLX engine finally gets a second model

The timing matters because of where Ollama is putting the model. Its MLX engine, the Apple Silicon path at the center of a Mac backend rewrite, launched as a preview narrowed to a single model: Qwen3.5-35B-A3B. It is a 35B parameter MoE that activates 3B per token, tuned for coding tasks and runnable on M-series Macs with more than 32GB of unified memory. That release said Ollama was "actively working to support future models." Muse Glimmer is the first real sign that expansion is underway, and it lands at a moment when Alibaba is also preparing to publish its largest model's weights.

The engine was upgraded along the way. The MLX update brought NVFP4 quantization for higher quality 4-bit models, a 20% inference speedup through fused Metal kernels, and snapshot caching that stops multi-agent, thinking-model, and branching workloads from reprocessing shared context. Those are the workloads Glimmer is being pointed at: coding agents that stay resident instead of calling out to the cloud.

The agent hook is the point. Recent Ollama work pushed tool calling through a single launch command, so coding agents such as Claude Code can drive a local model directly. Glimmer slots into that pipeline: open weights in, always-on local agent out. For a Mac user with enough memory, the path is the standard ollama run flow rather than a cloud dashboard.

What the two posts leave out

The gaps are as notable as the timing. The state-of-the-art claim on Apple Silicon has no benchmark numbers attached, and neither post says what hardware Glimmer needs, what its context window is, or when exactly Muse Spark 1.2, the model behind Meta's Muse Code terminal agent, lands. Ollama's post trails off with "Additional support and…" before the thread cuts out, so support beyond MLX is, for now, an ellipsis.

What the two posts establish, side by side:

Why same-day support is the real announcement

Same-day support carries weight for both sides. Meta gets an open-weights release in front of developers where they already run local models, with the tooling in place. Ollama gets a fresh flagship open model as proof its Apple engine has moved past the single-model preview. Both are trading on the same fact: the model works today, not at some announced future date.

Whether a 30B dense model is enough for the always-on agent use cases Ollama is pitching remains an open question, and tool use is not a free win: a recent paper finds tool-using models can make easy questions harder. Glimmer is a different architecture from the sparse MoE of the earlier preview model, and whether its 30B dense weights hold up for agent workloads is exactly what the next few weeks of community testing will determine. What the release settles is narrower but worth stating: a 30B model from a major lab, running natively on a Mac the day its weights went public. Whatever ships next will be measured against that.