Meta's 30B Muse Glimmer hits Apple Silicon first, NVIDIA support follows

Open weights are half of a release. The other half is a runtime, and Meta and Ollama delivered both on the same day. Meta Superintelligence Labs opened the weights for Muse Glimmer, its first release, and Ollama made it available through its MLX engine on Apple Silicon immediately.

The launch order is worth pausing on. Ollama's initial support covers Apple Silicon only, with NVIDIA, AMD, and other platforms promised in the coming days. An open model whose first official runtime is Apple Silicon rather than NVIDIA is the opposite of the usual order, and it changes what Mac owners can do today: run a 30B agent model locally, no cloud account required.

A 30B multimodal model under Apache 2.0

Muse Glimmer's spec sheet is aimed squarely at agent workloads. It is a 30B multimodal model with a 128K+ context length, released under the Apache 2.0 license. Reasoning strength is adjustable between low, medium, high, and xhigh; Ollama suggests high or xhigh for complex coding and agentic tasks, and lower settings when speed matters more.

Claude Code, Codex, and the local agent crowd

The model plugs into the agent ecosystem rather than sitting in a notebook. Ollama lists Claude Code, Codex, and Pi among the coding agents Muse Glimmer can power, plus OpenClaw and Hermes for long-running personal assistants. The ollama launch command also covers OpenCode and GitHub Copilot.

Setup is one command: ollama run muse-glimmer:30b-mlx. To run it under Claude Code, it is ollama launch claude, model muse-glimmer:30b-mlx. The thinking behind that syntax is Ollama's usual pitch: agents that stay resident on the machine instead of calling out to the cloud.

Apple Silicon gets DFlash before NVIDIA gets anything

Ollama's MLX engine now supports DFlash, building on its existing multi-token prediction support, and the company says Muse Glimmer runs 1.5x to 1.8x faster on Apple Silicon because of it. That is what makes back-to-back tool calls on a laptop plausible.

The Apple-first delivery is the story's twist, but it does not mean the model is Mac-bound. As we reported when the weights landed, a DFlash drafter has delivered up to 3.1x faster decoding on an RTX 5090, and quantization keeps the model under 20GB. The NVIDIA and AMD versions of the Ollama runtime are simply not out yet.

None of this came out of nowhere. Recent work on the MLX engine added NVFP4 quantization for higher-quality 4-bit models, a 20% inference speedup through fused Metal kernels, and snapshot caching that stops multi-agent and branching workloads from reprocessing shared context. Those are exactly the workloads Glimmer is pointed at: resident coding agents, long-running personal assistants, screenshot-driven computer use.

Meta's open source bet, one model at a time

Zuckerberg introduced the release on X with a description that matches the spec sheet: a "great 30B parameter dense model that can run locally." He also said Meta will open the weights for Muse Spark 1.2, its latest foundation model and the engine behind Muse Code, shortly. "Meta is a strong supporter of open source and I'm proud of these releases," he wrote.

The release lands as Ollama widens what it will run locally. A 35B coding-tuned model recently gained support on M-series Macs, and the company said it was actively working to support future models. Muse Glimmer is the first visible payoff of that work. Alibaba, meanwhile, is preparing to publish its largest model's weights, a move expected next week. The open-weights lane is filling up.

A single release is not a strategy. The pattern here is consistent though: open weights, a same-day local runtime, Apple Silicon first, and a next model already teased. That is how an open source bet becomes a habit, the same bet Meta has funded with billions.