Muse Glimmer: Meta’s Open-Weight AI Model

What Meta released

• Muse Glimmer is a 30-billion-parameter model released under the Apache 2.0 license.
• It is distilled from Meta’s larger, closed Muse Spark model and is designed to run locally on consumer hardware.
• Meta says its performance compares favorably with Gemma 4 and is competitive with Qwen 3.6.
• On a prompt-injection benchmark, attacks reportedly succeeded 28% of the time—better than Qwen’s 40%, but still a notable security risk.

How it runs on consumer GPUs

• Quantization compresses the model’s weights to roughly 4-bit precision, reducing memory use from over 55 GB to just under 20 GB.
• Speculative decoding uses a smaller model, DFlash, to propose blocks of tokens; Muse Glimmer checks them in one pass. Meta reports up to a 3× speedup on an RTX 5090.

Why the release matters

• Meta had previously shifted away from open models after Llama 4 drew criticism over benchmark claims and the quality of downloadable models.
• The company invested heavily in AI talent and launched the closed Muse Spark model, but faced questions about its spending and competitiveness.
• The video presents Glimmer’s open release as a possible return to Meta’s open-weight strategy, while questioning whether the move reflects genuine commitment or business pressure.

Meta’s broader claims and plans

• Mark Zuckerberg argues that AI’s main risk is concentrated control by a small number of companies, rather than rogue superintelligence.
• He has called for frontier AI labs to share mid-training checkpoints with the U.S. government and announced a fund for towns hosting Meta data centers.
• Meta leaders say open weights for Muse Spark 1.2 may follow, potentially enabling people to self-host the model behind Meta’s coding agent.

Bottom line

Muse Glimmer is a capable, locally runnable model with a permissive license. Whatever Meta’s motivations, its release makes the model available for others to use and build on.