[{"data":1,"prerenderedAt":30},["ShallowReactive",2],{"nr-en-meta-muse-glimmer-open-weights-30b":3},{"slug":4,"title":5,"dek":6,"date":7,"time":8,"publishedAt":9,"updated":10,"updatedAt":10,"dateFmt":11,"updatedFmt":10,"kind":12,"tier":13,"author":14,"authorName":15,"topics":16,"tracker":22,"trackerLabel":23,"headlineStat":24,"image":25,"ogImage":26,"imageAlt":5,"csv":10,"minutes":27,"words":28,"html":29},"meta-muse-glimmer-open-weights-30b","Meta Releases Muse Glimmer – First Open 30B Agent Model","After more than a year's pause, Meta is releasing model weights again. The new Muse Glimmer agent model runs entirely locally on consumer hardware – and Zuckerberg uses the occasion to take a sharp shot at OpenAI and Anthropic.","2026-08-10","16:05","2026-08-10T16:05:00+02:00","","August 10, 2026","news","standard","ideal-syka","Ideal Syka",[17,18,19,20,21],"Open-Source AI","Agent Models","Meta","Local AI","AI Competition","\u002Fstand-der-ki","AI Progress","30 billion parameters, under 20 GB storage","\u002Fnewsroom\u002Fimg\u002Fmeta-muse-glimmer-open-weights-30b.webp","\u002Fog-nr\u002Fmeta-muse-glimmer-open-weights-30b.en.png",3,565,"\u003Cp>Meta has released \u003Cstrong>Muse Glimmer\u003C\u002Fstrong> – an open-source AI model with \u003Cstrong>30 billion parameters\u003C\u002Fstrong> available under the Apache 2.0 license on Hugging Face. It is the first open model from Meta's new \u003Cstrong>Superintelligence Labs\u003C\u002Fstrong> since Llama 4 in spring 2025 and marks Meta's return to the open-source arena after a turbulent period.\u003C\u002Fp>\n\u003Ch2>Key Facts\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Cstrong>Muse Glimmer\u003C\u002Fstrong>: 30B-parameter model, quantized to under \u003Cstrong>20 GB\u003C\u002Fstrong>, runs locally on Mac\u002FPC with a single consumer GPU\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Use case\u003C\u002Fstrong>: AI agents for tasks like calendar management, file handling, and web search – entirely on your own device\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Competition\u003C\u002Fstrong>: Glimmer leads in most benchmarks against Google's \u003Cstrong>Gemma4\u003C\u002Fstrong> and Alibaba's \u003Cstrong>Qwen3.6-27B\u003C\u002Fstrong>, but falls behind on desktop control and terminal tasks\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Announcement\u003C\u002Fstrong>: According to the Wall Street Journal, an open version of \u003Cstrong>Muse Spark 1.2\u003C\u002Fstrong>, Meta's currently strongest model, will follow in the coming weeks\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2>Local Over Cloud – The Hardware Strategy\u003C\u002Fh2>\n\u003Cp>Glimmer's core promise is privacy through local execution. At full precision, the model would require over \u003Cstrong>55 GB of storage\u003C\u002Fstrong>. Meta compresses the weights to \u003Cstrong>4-bit\u003C\u002Fstrong> and brings the model under 20 gigabytes – small enough for current consumer graphics cards and MacBooks. An additional helper model accelerates text output by up to \u003Cstrong>3.1x\u003C\u002Fstrong> according to Meta.\u003C\u002Fp>\n\u003Cp>The idea is straightforward: an agent managing calendars, files, and personal messages should run entirely locally without data flowing to the cloud. This distinguishes Glimmer from most commercial competitors.\u003C\u002Fp>\n\u003Ch2>How Good Is Glimmer Really?\u003C\u002Fh2>\n\u003Cp>Meta compares Glimmer with the two leading open models in its size class:\u003C\u002Fp>\n\u003Cdiv class=\"tbl-scroll\">\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Model\u003C\u002Fth>\n\u003Cth>Maker\u003C\u002Fth>\n\u003Cth>Strengths\u003C\u002Fth>\n\u003Cth>Weaknesses\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd>\u003Cstrong>Muse Glimmer\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd>Meta\u003C\u002Ftd>\n\u003Ctd>Agentic tasks, tool use, long contexts\u003C\u002Ftd>\n\u003Ctd>–\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Cstrong>Gemma4\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd>Google\u003C\u002Ftd>\n\u003Ctd>–\u003C\u002Ftd>\n\u003Ctd>Behind Glimmer in most benchmarks\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>\u003Cstrong>Qwen3.6-27B\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd>Alibaba\u003C\u002Ftd>\n\u003Ctd>Desktop control, terminal tasks\u003C\u002Ftd>\n\u003Ctd>Behind Glimmer on agentic tasks\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Cp>Bottom line: Meta is competitive again in the small open model space, but not dominant. Important caveat: Meta conducted most comparison benchmarks itself and admits its test environment was not optimized for competing models. Proceed with caution on vendor-supplied benchmarks.\u003C\u002Fp>\n\u003Ch2>Zuckerberg's Counterattack Against OpenAI and Anthropic\u003C\u002Fh2>\n\u003Cp>With the release, Zuckerberg published the essay \u003Cstrong>"The Future is for Everyone,"\u003C\u002Fstrong> laying out the rationale for Meta's open-source approach. The text reads like a direct response to OpenAI and Anthropic: Zuckerberg aggressively defends \u003Cstrong>distillation of competing models\u003C\u002Fstrong> – training smaller models to replicate the outputs of larger ones. This is the standard way most compact models are built today.\u003C\u002Fp>\n\u003Cp>Glimmer itself was trained via distillation from Meta's larger Muse Spark model. But Zuckerberg also uses the release to call for fewer restrictions on US AI labs – a direct counter to regulatory criticism from OpenAI CEO Sam Altman and Anthropic.\u003C\u002Fp>\n\u003Ch2>What This Means for German Enterprises\u003C\u002Fh2>\n\u003Cp>For developers and companies in Germany, Glimmer could become relevant where local data processing is a requirement – such as in administration, healthcare, or sensitive business processes. The open license enables customization without dependence on US cloud providers. However: the competition is not sleeping, and whether Glimmer prevails against specialized solutions depends on practical performance, not just benchmarks. Additionally, it remains unclear how stable Meta's open-source strategy is long-term, given that Llama 4 disappointed.\u003C\u002Fp>\n\u003Ch2>Sources\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fthe-decoder.de\u002Fmetas-neues-offenes-ki-modell-erscheint-mit-plaedoyer-fuer-das-lernen-aus-konkurrenz-daten-und-einer-600-milliarden-frage\u002F\">The Decoder (DE)\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fwww.foxbusiness.com\u002Fvideo\u002F6403152116112\">Fox Business\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fwww.benzinga.com\u002Ftrading-ideas\u002Flong-ideas\u002F26\u002F08\u002F61073538\u002Fmark-zuckerberg-145-billion-meta-bet-getting-serious\">Benzinga\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cem>Editorially owned by \u003Ca href=\"\u002Fen\u002Fautor\u002Fideal-syka\">Ideal Syka\u003C\u002Fa>. Sources and method: \u003Ca href=\"\u002Fen\u002Fredaktion\">Newsroom & method\u003C\u002Fa>. Tips and corrections: \u003Ca href=\"mailto:ai@i6eal.de\">ai@i6eal.de\u003C\u002Fa>.\u003C\u002Fem>\u003C\u002Fp>\n",1786378401162]