Meta open-sources 30B Muse Glimmer under Apache 2.0
Meta Superintelligence Labs released its first open model — a 30B agentic multimodal system under Apache 2.0 that runs on a single consumer GPU.
Meta published open weights for Muse Glimmer on Monday, a 30-billion-parameter multimodal model tuned for agent workloads and licensed under Apache 2.0. It is the first open release from Meta Superintelligence Labs, the unit the company assembled last year, and it arrives alongside a commitment from Mark Zuckerberg to publish weights for Muse Spark 1.2 — Meta's most capable foundation model — within the coming weeks.
What shipped
The model pairs a 2B ViT-style perception encoder with a 28B text decoder. Meta describes a hybrid attention stack that alternates sliding-window and full-attention layers, plus gated grouped-query attention that the company says cuts KV-cache memory by roughly 16x. Quantised to 4 bits the checkpoint lands under 20 GB, which puts it inside a 24 GB VRAM envelope; Meta reports testing on an RTX 5090 and on M4-Max and M5-Max Macs. An optional speculative-decoding drafter called DFlash raises generation speed by about 3.1x on the RTX 5090 and 1.5–1.8x on the Apple silicon configurations.
The pitch is agentic reliability at a size that fits on a desk. Muse Glimmer plans, calls tools, checks its own output and recovers from failures, accepts interleaved text, image and video input, and covers more than 100 languages. On Meta's published comparisons against Gemma4-31B and Qwen3.6-27B, it scores 75.5 on MCP Atlas (versus 62.5 and 54.2), 51.2 on SWE-Bench Pro (versus 50.2 and 36.9) and 76.0 on SWE-Bench Verified, where Qwen3.6-27B edges it at 77.2.
Weights are on Hugging Face with day-zero support across llama.cpp, vLLM, MLX, ExecuTorch, Ollama and LM Studio, plus hosted endpoints from Together AI and Fireworks AI.
Why it matters
Meta's retreat from open weights over the past year left the open-model tier to Chinese labs, which now account for a large share of Hugging Face downloads and set the pace on licensing. A permissively licensed, agent-tuned 30B model from Meta puts a well-resourced US lab back into that competition at exactly the size class developers actually deploy locally. The bigger variable is Spark 1.2: if Meta follows through on frontier-tier open weights, it would be the first time a US lab has released anything at that level since the Llama era, and it would reset what "open" means at the top of the market.