{"slug": "meta-releases-muse-spark-open-says-it-is-most-capable-us-based-open-model", "title": "Meta Releases Muse Spark Open, Says It Is Most Capable US-Based Open Model", "summary": "Meta released Muse Spark Open, an open-weight model it calls the most capable US-origin open model available, with weights live on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Meta Chief AI Officer Alexandr Wang said the model has 816 billion total parameters, 42 billion active parameters, a 1 million token context window, and native multimodality, and is purpose-built for coding and agentic work. Meta says the model can run locally on two linked Apple M5 Ultra Mac Studios via Apple's MLX framework, and ships under the Muse Model License with limited commercial terms allowing most developers to use it for free.", "body_md": "Meta has released Muse Spark Open, which the company is calling the most capable US-origin open-weight model available today. The model is live for download now, with weights hosted on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.\n\nAlexandr Wang, [Meta’s Chief AI Officer](https://officechai.com/ai/former-scale-ai-ceo-alexandr-wang-becomes-chief-ai-officer-at-meta/), announced the release on X, writing that Muse Spark Open is “the most capable US open-weight model to date,” built with 816 billion total parameters and 42 billion active parameters, a 1 million token context window, and native multimodality. He described it as purpose-built for coding and agentic work, including long-horizon tasks, complex software engineering, and multi-step reasoning, and capped the thread off with “america!”\n\nThe release marks a shift back toward open weights for Meta’s AI division. Muse Spark, the closed, proprietary flagship model that powers the Meta AI assistant, launched earlier this year as the first product out of Meta Superintelligence Labs. Muse Spark Open takes the same naming convention but goes the opposite direction on distribution, with Meta releasing the full weights for anyone to download and run.\n\n## Built to Run on Your Desktop\n\nMeta is positioning Muse Spark Open as the first model in its class that can run locally on consumer-grade hardware, specifically two Apple M5 Ultra Mac Studios linked together using Apple’s MLX framework. The company says the M5 Ultra’s unified memory architecture, which allows its integrated GPU to address up to 512GB per machine, lets two Mac Studios combine to behave like one machine with a terabyte of unified memory, enough to hold the entire model.\n\nMeta says it tested the setup on real agentic workflows running entirely on-device, including shipping iOS changes to the open-source Mastodon client Ice Cubes, with the model handling bug fixes, new features, and a full build, and driving DaVinci Resolve end-to-end, reviewing raw footage to identify and crop the strongest moments.\n\n## Benchmarks and Availability\n\nOn benchmarks shared by Meta, Muse Spark Open leads rival open models including Nemotron 3 Ultra, Mistral Large 4, Qwen3.8, GLM 5.3, and Kimi K3 on agentic evaluations like MCP-Atlas and JobBench, and on reasoning tests such as HLE-Diamond and AA-Omniscience. It trails some closed Chinese models like Kimi K3 and GLM 5.3 on specific coding benchmarks such as Terminal-Bench 2.1, but Meta is pitching it as the strongest option built and released in the US.\n\nBeyond the model-weight downloads, Muse Spark Open is also live through API platforms Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. It’s designed to run across inference engines including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and Nvidia silicon. The open-weight release also ships with custom Helion GPU kernels in SGLang, which Meta says deliver 1.03-3.95x kernel-level speedups on Nvidia and AMD GPUs.\n\nDevelopers looking to customize the model further can do so using the PyTorch framework, including PyTorch’s TorchTitan for additional tuning. Meta says it evaluated the model against its internal Superintelligence Scaling Framework covering the most severe risk categories before release, and that Muse Spark Open ships with additional safety post-training given the considerations that come with distributing open weights. The weights themselves are released under the Muse Model License, which Meta says carries limited commercial terms allowing most developers to use the model for free.", "url": "https://wpnews.pro/news/meta-releases-muse-spark-open-says-it-is-most-capable-us-based-open-model", "canonical_source": "https://officechai.com/ai/muse-spark-open-benchmarks/", "published_at": "2026-10-15 19:13:43+00:00", "updated_at": "2026-10-15 20:26:57.348405+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-products", "ai-tools", "ai-agents"], "entities": ["Meta", "Muse Spark Open", "Alexandr Wang", "Hugging Face", "Amazon SageMaker AI", "Google Cloud Model Garden", "Microsoft Foundry Models", "Apple MLX"], "also_reported_by": [{"url": "https://www.thedeepview.com/articles/meta-makes-its-open-models-comeback", "source": "thedeepview.com", "title": "Meta makes its open models comeback", "published_at": "2026-10-15T19:44:39+00:00"}], "alternates": {"html": "https://wpnews.pro/news/meta-releases-muse-spark-open-says-it-is-most-capable-us-based-open-model", "markdown": "https://wpnews.pro/news/meta-releases-muse-spark-open-says-it-is-most-capable-us-based-open-model.md", "text": "https://wpnews.pro/news/meta-releases-muse-spark-open-says-it-is-most-capable-us-based-open-model.txt", "jsonld": "https://wpnews.pro/news/meta-releases-muse-spark-open-says-it-is-most-capable-us-based-open-model.jsonld"}}