Meta's Muse Glimmer: a 30B Apache-2.0 agent model runs on one GPU
Meta's back in open source: a 131K-context model that quantizes into 24GB and beats Qwen and Gemma on tool-use benches — plus MiniMax-H3 now runs on Macs.
Copy markdown
24GB VRAM, 131K context, Apache 2.0
The 30B dense model 4-bit quantizes onto one 24GB GPU (or an M4/M5 Max) with ~1% accuracy loss, at a 131,072-token context. Weights ship on Hugging Face as BF16, GGUF k-quants and ExecuTorch builds — commercial use included.
It tops Qwen and Gemma on agent benches
Muse Glimmer hits 75.5 on MCP Atlas (Gemma4-31B 54.2, Qwen3.6-27B 62.5), 51.2 on SWE-Bench Pro and 94.7 on AIME 2026. It trails Qwen3.6-27B on GUI control (OSWorld 65.9 vs 75.6), so reach for it on tool-use and reasoning, not desktop clicking.
Build this weekend: a local screenshot-reading agent
It's tuned for desktop agents that read screenshots plus schema-based function calling, so you can wire a fully offline agent that drives your own apps. An RTX 5090 gets a 3.1x decode speedup from the bundled DFlash drafter.
MiniMax-H3 now makes video on your Mac
antirez shipped h3.c, a native engine running the 33B MiniMax-H3 video-plus-audio model on an M5 Pro/Max — about 40GB of memory and roughly an hour for a 480x864 clip, no cloud. Heads up: H3's weights reportedly bar US/EU use, so read the license first.