Meta releases open-weight Muse Glimmer, unwinds $2B Manus AI deal

Muse Glimmer marks Meta's return toward open weights after a period of retrenchment — a 30B Apache 2.0-licensed agentic model built to run locally on consumer GPUs via SGLang, vLLM, or NVIDIA NIM, keeping data off the cloud. Crucially, Meta is hedging: the more capable Muse Spark stays proprietary, so this is a two-tier strategy, not a wholesale open-source reconversion. The release came wrapped in Zuckerberg's 6,500-word 'AI for everyone' letter, an ideological framing meant to reclaim open-source credibility.

The Atlantic's Nicholas Thompson captured the strategic logic on LinkedIn (342 likes): Meta's apparent yo-yoing on open source actually makes business sense because Meta doesn't need to lead the frontier — it needs an AI that's 'good enough' to power its existing products while keeping competition among everyone else tight. Open-weighting a mid-tier model commoditizes rivals' offerings without giving away Meta's best work.

The second development is corporate: Meta will fully unwind its $2 billion acquisition of Chinese platform Manus after Beijing blocked the deal — a concrete casualty of US-China tech decoupling. Competitively, Muse Glimmer lands into a market where Qwen already dominates local inference (3B downloads) and NVIDIA's Nemotron 3.5 Lightning targets the same on-device agent niche, so a 30B open release faces stiff mindshare competition. Watch whether developers adopt Glimmer or treat Spark's proprietary status as a reason to stay on Qwen.