{"slug": "meta-releases-muse-glimmer-weights-and-plans-muse-spark-1-2-release", "title": "Meta releases Muse Glimmer weights and plans Muse Spark 1.2 release", "summary": "Meta released the weights for Muse Glimmer, a 30-billion-parameter multimodal model under the Apache 2.0 license, aimed at local agentic and coding workloads, and said it plans to publish weights for Muse Spark 1.2 in the coming weeks. The model supports a context window of more than 131,000 tokens and can run locally after quantization, with a 4-bit deployment envelope of roughly 20 gigabytes. Meta CEO Mark Zuckerberg paired the announcement with a defense of broadly available AI systems and fewer barriers to U.S. AI development.", "body_md": "# Meta releases Muse Glimmer weights and plans Muse Spark 1.2 release\n\n- Meta released Muse Glimmer, a 30-billion-parameter dense multimodal model aimed at local agentic and coding workloads.\n[[1]](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) - Meta says the model supports a context window of more than 131,000 tokens and can run locally after quantization, depending on the hardware and runtime.\n[[1]](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) - The company said it plans to release weights for Muse Spark 1.2 in the coming weeks, but has not yet published a final release date or license for that model.\n[[2]](https://www.cnbc.com/2026/08/10/meta-muse-glimmer-open-weight-ai.html) - Mark Zuckerberg paired the announcement with a defense of broadly available AI systems and an argument for fewer barriers to U.S. AI development.\n[[3]](https://www.axios.com/2026/08/10/zuckerberg-ai-manifesto-meta)\n\nMeta on Monday released the weights for Muse Glimmer, a 30-billion-parameter multimodal model designed to run locally, and said it plans to publish the weights for Muse Spark 1.2 in the coming weeks. The announcement reopens Meta’s participation in the open-weight model market after its new Muse family initially launched with proprietary models. [[1]](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model)[[2]](https://www.cnbc.com/2026/08/10/meta-muse-glimmer-open-weight-ai.html)\n\nMuse Glimmer is aimed at agentic and coding workloads rather than positioned as Meta’s primary consumer assistant. Meta describes it as a dense model with support for tool use, long-context tasks and more than 100 languages. Its published materials list a context window above 131,000 tokens and describe a quantized configuration intended for local deployment. [[1]](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model)\n\n## A local model for agents\n\nMeta says Muse Glimmer is released under the Apache 2.0 license and is designed for developers who want to run agents on their own hardware instead of sending every prompt and tool call to a cloud provider. The model is multimodal and tuned for coding, tool use and other tasks that require a system to take multiple steps. [[1]](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model)\n\nThe local-hardware claim depends on configuration. Quantization reduces memory requirements, but total usage also varies with precision, context length, the vision component, runtime overhead and any speculative-decoding components. Meta’s materials describe a 4-bit deployment envelope of roughly 20 gigabytes, but that should not be read as a universal requirement for every use case. [[1]](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model)\n\nMeta’s benchmark results are company-reported and should be read with the evaluation setup in mind. Results for coding and tool-use models can change substantially with the harness, system prompt, tool definitions, sampling settings and whether the measurement covers the model alone or the complete agent system. [[1]](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model)\n\n## Spark 1.2 is not available yet\n\nThe planned Muse Spark 1.2 release is the more important strategic signal. Muse Spark launched on April 8 as Meta Superintelligence Labs’ first model and now powers Meta AI across the company’s consumer products. Meta introduced Spark 1.1 in July with features for planning, connecting to applications and completing multi-step tasks. [[4]](https://about.fb.com/news/2026/04/introducing-muse-spark-meta-superintelligence-labs/)[[5]](https://about.fb.com/news/2026/07/meta-ai-muse-spark-doesnt-just-think-it-acts/)\n\nMeta has not yet published the Spark 1.2 weights, a final release date, a complete license or the hardware specifications needed to run them. The announcement therefore establishes a planned release, not a completed open-weight launch. [[2]](https://www.cnbc.com/2026/08/10/meta-muse-glimmer-open-weight-ai.html)\n\nThe distinction between open weights and open source also remains material. Publishing weights can permit local inference, fine-tuning and independent evaluation, but it does not necessarily include training data, training code, evaluation infrastructure or unrestricted use. The eventual Spark 1.2 license and model card will determine what developers can do with the release. [[1]](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model)[[2]](https://www.cnbc.com/2026/08/10/meta-muse-glimmer-open-weight-ai.html)\n\n## Zuckerberg makes the policy case\n\nMark Zuckerberg paired the model announcement with a 6,500-word essay arguing that one of the main risks of advanced AI is excessive control by a single company or institution. He used that argument to defend broadly available models and Meta’s decision to resume releasing some systems with open weights. [[3]](https://www.axios.com/2026/08/10/zuckerberg-ai-manifesto-meta)\n\nZuckerberg also argued for policies that preserve room for U.S. AI development and criticized restrictions he said could slow American companies. His position aligns Meta with a broader technology-industry campaign against sweeping limits on open models, although the policy debate has not settled how safeguards should apply to models that can execute code, access files or call external services. [[3]](https://www.axios.com/2026/08/10/zuckerberg-ai-manifesto-meta)[[6]](https://www.washingtonpost.com/technology/2026/07/24/top-tech-firms-urge-us-government-not-limit-open-ai-models/)\n\nMeta’s release materials still include safety and use restrictions, and a locally run model does not remove deployment risks. Developers remain responsible for access controls, monitoring and the permissions granted to an agent. Nothing in the announcement establishes that Muse Glimmer is suitable for unsupervised high-impact decisions. [[1]](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model)\n\n## Companies mentioned\n\n## Further sources\n\n[[1] Meta AI Research, “Introducing Muse Glimmer: An Open Agentic Model That Runs on… ↗](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model)\n\n[[2] CNBC, “Meta to open source its most powerful AI model as it takes swipe at Open… ↗](https://www.cnbc.com/2026/08/10/meta-muse-glimmer-open-weight-ai.html)\n\n[[3] Axios, “Zuckerberg: AI's biggest risk is one entity with too much control,” Aug… ↗](https://www.axios.com/2026/08/10/zuckerberg-ai-manifesto-meta)\n\n[[4] Meta, “Introducing Muse Spark: Meta’s Most Powerful Model Yet,” April 8, 2026, … ↗](https://about.fb.com/news/2026/04/introducing-muse-spark-meta-superintelligence-labs/)\n\n[[5] Meta, “Meta AI Doesn’t Just Think, It Acts,” July 24, 2026. ↗](https://about.fb.com/news/2026/07/meta-ai-muse-spark-doesnt-just-think-it-acts/)\n\n[[6] The Washington Post, “Top tech firms urge U.S. government not to limit ‘open’ A… ↗](https://www.washingtonpost.com/technology/2026/07/24/top-tech-firms-urge-us-government-not-limit-open-ai-models/)\n\nThe stories that matter, in one email. Free — unsubscribe anytime.", "url": "https://wpnews.pro/news/meta-releases-muse-glimmer-weights-and-plans-muse-spark-1-2-release", "canonical_source": "https://mlq.ai/news/meta-releases-muse-glimmer-weights-and-plans-muse-spark-12-release/", "published_at": "2026-08-10 17:34:32.479697+00:00", "updated_at": "2026-08-10 17:34:34.496181+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-products", "ai-policy"], "entities": ["Meta", "Muse Glimmer", "Muse Spark 1.2", "Mark Zuckerberg", "Meta Superintelligence Labs", "Apache 2.0"], "alternates": {"html": "https://wpnews.pro/news/meta-releases-muse-glimmer-weights-and-plans-muse-spark-1-2-release", "markdown": "https://wpnews.pro/news/meta-releases-muse-glimmer-weights-and-plans-muse-spark-1-2-release.md", "text": "https://wpnews.pro/news/meta-releases-muse-glimmer-weights-and-plans-muse-spark-1-2-release.txt", "jsonld": "https://wpnews.pro/news/meta-releases-muse-glimmer-weights-and-plans-muse-spark-1-2-release.jsonld"}}