Meta launched Muse Glimmer on August 10-11, a 30-billion-parameter AI model with open weights to run on consumer hardware such as Macs and PCs with a single GPU [1, 2, 3, 4, 5, 6]. The model’s weights are licensed under Apache 2.0 allowing users to download, modify, and redistribute Muse Glimmer freely [2, 4, 7, 6].

Muse Glimmer is distilled from Meta’s larger closed Muse Spark model, first introduced in April 2026 and updated to version 1.2 on August 5 alongside Muse Code, a programming assistant powered by Muse Spark [1, 2, 3, 4, 6]. Unlike Muse Spark, which requires significant hardware resources, Muse Glimmer can run with less than 20GB of memory through quantization techniques, enabling local use on typical consumer systems [5].

The model supports multi-step workflows, multitasking with text and image inputs, and features reliability enhancements such as failure recovery. Muse Glimmer was trained on data from over 100 languages, emphasizing broad global utility [2, 4, 5, 6].

Meta CEO Mark Zuckerberg published a lengthy essay promoting open AI models to empower individuals rather than concentrate AI control in a few companies. He said, "Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it" [1, 2, 3, 4, 7, 8, 6]. Zuckerberg also noted US labs face restrictions on training data compared to foreign competitors and called for policy reforms to support American open source AI leadership [3].

Meta’s approach contrasts with companies like OpenAI and Anthropic, which focus on closed or cloud-only AI models and favor strict AI regulation [1, 2, 3, 7, 8, 6]. Industry voices welcomed Meta’s open-weight effort with comments like "Good move. Bravo" from Yann LeCun and praise from Aaron Levie stating, "America now finally has its response to the open weights AI race" [8].

Chinese AI startups have released powerful open-weight models, intensifying competition and policy debates about US AI innovation [1, 3]. While Muse Glimmer and Muse Spark lag some top models on benchmarks, they are competitive in cost and openness [1, 2, 3].

Looking ahead, Meta plans to release a more powerful AI model internally codenamed "Watermelon" though its openness remains unclear [1].