❯ Meta Releases Open-Source Muse Glimmer: 30B-Parameter Model Runs on a Single Consumer GPU

OPEN SOURCEMeta released Muse Glimmer, a 30B-parameter dense multimodal model under an Apache 2.0 license. After 4-bit quantization it compresses to under 20GB, allowing persistent agents to run on a single consumer-grade GPU — Meta’s first genuinely open-source license since the Llama series (which used a custom, non-OSI-certified license). Meta also said the stronger Muse Spark 1.2 weights will follow as open source “within the coming weeks.”

TRAININGGlimmer is not a base model trained from scratch — it was distilled directly from Muse Spark via logits, trained for agentic tasks from the start. It supports a 131K context window and 100+ languages, and achieves a 3.1x speedup on the RTX 5090 via speculative decoding.

BENCHMARK GAPActual measurements from third-party benchmark firm Artificial Analysis show it still trails Qwen3.6 27B and Gemini 3.5 Flash-Lite — models from the same cycle — on agentic evals. Researcher teortaxesTex cautions that the Qwen and Gemma models used for comparison are both April releases, so a generation gap between the two sides already existed from the outset.

THE MONEYThe release is paired with a 6,500-character Zuckerberg essay; Meta’s official announcement says it will set up a $1 billion community fund to compensate areas around data centers — against the backdrop of Meta’s capital expenditure projected to surge to $145 billion this year. What the open-source camp gains this time is the “model that can actually do work on consumer hardware” slot, but whether the gap to China’s frontier open-source models has truly narrowed won’t be verifiable until the Spark 1.2 weights land.

▪ SIGNALWhat Meta is filling this time isn’t a model-capability gap — it’s the “is the open-source license clean?” question the company itself once botched.