Meta released Muse Glimmer, a 30-billion-parameter dense multimodal model, under an Apache 2.0 licence on 10 August. The weights went up on Hugging Face; there is no post on Meta's AI blog, whose newest entries are still from late July.
The licence is the news
Meta's previous open releases carried bespoke terms with acceptable-use and scale restrictions. Apache 2.0 has none of that: commercial use, modification and redistribution are permitted outright. That closes the specific gap that pushed developers toward Qwen and Gemma.
It fits on one card
At 4-bit quantisation the model compresses to roughly 20GB, inside the budget of a 24-32GB consumer GPU. Meta pairs it with DFlash speculative decoding, claiming 3.1x faster generation on an RTX 5090 and 1.8x on an Apple M5 Max. It covers over 100 languages and is benchmarked on agentic suites including SWE-Bench and MCP-Atlas, positioned against Gemma4-31B and Qwen3.6-27B.
Distilled from a model nobody can download
Glimmer is distilled from Muse Spark 1.2, Meta's closed flagship. Accounts differ on whether Spark 1.2's own weights are coming: one reading has it shipping as a closed model on 5 August with weights promised later. Meta has not published a schedule.
Released sideways
The absence of a newsroom post matters for verification: the primary artefacts are a model repository and a roughly 6,500-word Zuckerberg essay on distributed AI. Benchmark claims are Meta's and are not independently reproduced.
The slot it is fighting for
The 30B class is where self-hosting decisions get made: large enough for agentic work, small enough to run without renting a cluster. Qwen and Gemma have held that slot for a year, partly on licence terms. Matching Apache 2.0 removes the legal objection and leaves the argument to benchmarks nobody outside Meta has reproduced.