Meta Returns to Open Source: Releases Apache 2.0 Model MuseGlimmer, Zuckerberg Advocates for Individual AI Empowerment

机器之心 · wechat · 2026-08-11

Meta is embracing the open-weight approach again with the release of MuseGlimmer. The model features 29.6B parameters and uses the highly permissive Apache 2.0 license, allowing direct commercial deployment and modification.

Capabilities & Positioning

Distilled from the more powerful MuseSpark, MuseGlimmer supports text/image input with a 128K context. It focuses on multi-step reasoning, tool calling, and fault recovery, positioned as an on-device Agent foundation.
Hardware requirements: BF16 weights are 60GB, but a 4-bit quantized version fits into 20GB, runnable on 24GB/32GB VRAM setups or high-end Macs.
Evaluations: Shows competitive intelligence and leads in tool-calling agent benchmarks, but suffers from a high hallucination rate in knowledge tasks, requiring retrieval augmentation.

Zuckerberg's AI Power Narrative

Zuckerberg published an essay emphasizing the need to provide free, powerful AI to billions to counterbalance large institutions.
He argues against restricting model access for safety reasons and defends model distillation, asserting that open models are ultimately safer and more transparent.

Ecosystem & Compute Strategy

Facing fierce competition from Chinese open-source models like DeepSeek, Kimi, and Qwen, Meta aims to retain global developers.
This open strategy complements Meta's massive $145B AI infrastructure investment. While weights are free, Meta can monetize through inference, hosting, and cloud services. The more powerful MuseSpark 1.2 weights will be released in the coming weeks.

Original post →

Meta Returns to Open Source: Releases Apache 2.0 Model MuseGlimmer, Zuckerberg Advocates for Individual AI Empowerment

More from Companies & People