On August 10, Meta released Muse Glimmer under an Apache 2.0 license, marking a sharp return to the open-source strategy that made Llama famous. The 30-billion-parameter model is optimized for on-device agent workflows and fits into 18-20 GB of RAM after 4-bit quantization—small enough to run on a Mac or PC with a single consumer GPU.
Muse Glimmer outperformed similar-sized rivals like Gemma 4 and Qwen 3.6 on agentic, coding, and reasoning benchmarks. It scored 75.5 on MCP Atlas, 51.2 on SWE-Bench Pro, and 94.7 on AIME 2026. Meta Superintelligence Labs chief Alexandr Wang said the company would release the weights for Muse Spark 1.2—its frontier-class model—soon, instantly making it the top U.S. open rival to China's ecosystem.
Zuckerberg's 6,500-word case for open AI
Alongside the model, Mark Zuckerberg published an essay arguing that concentrated AI control is inherently problematic and that U.S. policy slowing open-source releases adds risk to American leadership. Meta also announced a 1 billion dollar fund for communities hosting its data centers and committed its board to approve future superintelligence safety criteria.
The timing matters: Meta paused open-weight releases after reorganizing its AI operations last year. This restart—with models that actually compete—signals the vibe shift is real.