Meta's Muse Glimmer: a 30B agentic model, Apache 2.0, runs on one GPU
Meta's return to open weights is tuned for local coding agents and downloadable now — while the frontier Muse Spark stays closed and MiniMax-H3 hits Macs.
Copy markdown
30B, Apache 2.0, fits one consumer GPU
Muse Glimmer is a 30B open-weight model tuned for tool use, long tasks, and failure recovery, sized to run on a single 24–32GB consumer GPU. Apache 2.0 means you can ship it commercially with no license strings.
Built for local agents, not chatbot demos
Meta trained it end-to-end for agentic task completion, multi-step reasoning, and LLM-as-judge eval — the always-on, on-device niche. Ollama shipped same-day support (0.32.7), so `ollama run` works right now.
Weights are live on Hugging Face today
Grab them at meta-models/Muse-Glimmer-30B under Apache 2.0, quantize, and wire it into your agent loop this weekend — no API key, no rate limits, and nothing leaving your machine.
The frontier stays closed: Muse Spark 1.2
Glimmer is the on-device sibling, not Meta's best model. The flagship Muse Spark 1.2 is still unreleased with licensing undecided, so temper expectations on raw capability.
Build this weekend: MiniMax-H3 video on your Mac
A fresh MLX port runs MiniMax-H3's 33B joint video-and-audio diffusion natively on Apple Silicon — no CUDA, no cloud. Generate narrated clips locally on an M-series Mac.