{"id":762,"date":"2026-10-16T13:43:25","date_gmt":"2026-10-16T13:43:25","guid":{"rendered":"https:\/\/managedt.com\/blog\/muse-spark-open-weights-released-816b-open-weight-model-mac-studio\/"},"modified":"2026-11-05T11:16:46","modified_gmt":"2026-11-05T11:16:46","slug":"muse-spark-open-weights-released-816b-open-weight-model-mac-studio","status":"publish","type":"post","link":"https:\/\/managedt.com\/blog\/muse-spark-open-weights-released-816b-open-weight-model-mac-studio\/","title":{"rendered":"Muse Spark Open Weights Released: An 816B Open-Weight Model That Runs on Two Mac Studios"},"content":{"rendered":"<p>Meta released Muse Spark Open on October 15, 2026. According to Meta&#8217;s <a href=\"https:\/\/research.meta.ai\/blog\/muse-spark-open\" target=\"_blank\" rel=\"noopener\">release post<\/a>, it is an 816-billion-parameter open-weight model that was optimized to run on two Apple M5 Ultra Mac Studios.<\/p>\n<div>\n<p><strong>What this means for a small-business IT setup<\/strong><\/p>\n<ul>\n<li>The local configuration described by Meta uses two Mac Studios, whose integrated GPUs can each use up to 512 GB of unified memory.<\/li>\n<li>Identify who will assemble the system, connect the machines, load the weights, monitor performance, apply updates and handle failures.<\/li>\n<li>The team must compare the cost and control of local deployment with hosted inference.<\/li>\n<li>Customer, employee and other sensitive information should not enter the system until storage, access and output-review policies are approved.<\/li>\n<\/ul>\n<\/div>\n<h2>What Is Muse Spark Open?<\/h2>\n<p>Muse Spark Open is Meta&#8217;s sparse mixture-of-experts model for coding and agentic work. It can accept text, image, video and audio input, while the official developer documentation lists text as its output. Its agentic design is intended for work that spans many steps, such as reading a codebase, planning a change, calling tools and checking the result.<\/p>\n<table>\n<thead>\n<tr>\n<th>Specification<\/th>\n<th>Stated detail<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Total parameters<\/td>\n<td>816 billion<\/td>\n<\/tr>\n<tr>\n<td>Active parameters<\/td>\n<td>Approximately 42 billion per token<\/td>\n<\/tr>\n<tr>\n<td>Context length<\/td>\n<td>1 million tokens<\/td>\n<\/tr>\n<tr>\n<td>Hardware configuration<\/td>\n<td>Two Mac Studios with M5 Ultra chips, whose integrated GPUs can each use up to 512 GB of unified memory<\/td>\n<\/tr>\n<tr>\n<td>License<\/td>\n<td>Muse Model License with limited commercial terms<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Meta released the model under the Muse Model License, which allows many developers to use it for free but includes commercial conditions. A separate licensing summary reports that organizations whose aggregate gross revenue with affiliates exceeds $20 million in any consecutive 12-month period must notify Meta and enter a separate written agreement before offering the model or derivatives as a Model-as-a-Service. Internal use, research and non-MaaS commercial product features are exempt from that threshold. See the <a href=\"https:\/\/llm-stats.com\/models\/muse-spark-open\" target=\"_blank\" rel=\"noopener\">Muse Spark Open license and pricing summary<\/a>.<\/p>\n<h2>What Does the Two-Mac-Studio Configuration Require?<\/h2>\n<p>Meta says the M5 Ultra integrated GPU in each Mac Studio can use up to 512 GB of unified memory, allowing two machines to hold the model. Each chip provides 1.2 TB\/s of memory bandwidth, which Meta says is sufficient to read the small portion of weights needed for each token. Two such machines provide one terabyte of unified memory in total.<\/p>\n<p>The practical decision is not simply whether the hardware can hold the model. An IT lead also needs to establish who will assemble the system, connect the machines, load the weights, monitor performance, apply updates and handle failures.<\/p>\n<p>Meta&#8217;s <a href=\"https:\/\/research.meta.ai\/blog\/muse-spark-open\" target=\"_blank\" rel=\"noopener\">October 15 release post<\/a> describes the optimized Apple silicon configuration. Its <a href=\"https:\/\/dev.meta.ai\/resources\/blog\/build-with-muse-spark-open\" target=\"_blank\" rel=\"noopener\">developer documentation<\/a> also lists support for SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth across NVIDIA, AMD, Intel and Apple hardware.<\/p>\n<h2>What Did Meta Demonstrate?<\/h2>\n<p>Meta tested agentic workflows that ran entirely on-device with Muse Code powered by Muse Spark Open. One example shipped iOS changes to Ice Cubes, an open-source Mastodon client, including bug fixes, new features and a complete build. The other drove DaVinci Resolve end to end, reviewing footage frame by frame to select strong moments and crop them vertically.<\/p>\n<p>The <a href=\"https:\/\/theairankings.com\/meta\/muse-spark-open\/\" target=\"_blank\" rel=\"noopener\">AI Rankings review<\/a> notes that Meta&#8217;s launch post printed no benchmark figures and that the model has no published SWE-bench score.<\/p>\n<p>Independent benchmark coverage remains mixed. BenchLM&#8217;s <a href=\"https:\/\/benchlm.ai\/models\/muse-spark-open\" target=\"_blank\" rel=\"noopener\">November 4 model profile<\/a> reports a 50.5 decision score and ranks the model 98th among 216 entries, while placing coding 43rd among 156 eligible models and agentic work 51st among 135. It also notes that eight benchmark rows have displayable evidence out of 797 tracked slots, so its public position does not yet have enough sourced coverage for a verified rank.<\/p>\n<p>Artificial Analysis figures reported by <a href=\"https:\/\/theairankings.com\/meta\/muse-spark-open\/\" target=\"_blank\" rel=\"noopener\">The AI Rankings<\/a> give the model an Intelligence Index v5.0 score of 25 at xhigh reasoning, an output speed of 139 tokens per second and a median hosted price of $1.25 per million input tokens and $4.25 per million output tokens. The same source notes that Meta&#8217;s launch post printed no benchmark figures and that the model has no published SWE-bench score.<\/p>\n<h2>How Does Hosted Access Compare?<\/h2>\n<p>Muse Spark Open is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Model weights can also be downloaded from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models.<\/p>\n<p>The team should review the provider&#8217;s data-handling terms, security controls, service levels and support before selecting a route.<\/p>\n<p>Hosted pricing information is not identical across public catalogs. The <a href=\"https:\/\/llm-stats.com\/models\/muse-spark-open\" target=\"_blank\" rel=\"noopener\">LLM Stats pricing page<\/a> lists Together AI at $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens. BenchLM&#8217;s <a href=\"https:\/\/benchlm.ai\/models\/muse-spark-open\" target=\"_blank\" rel=\"noopener\">November profile<\/a> says no comparable first-party hosted token rate is published in its catalog. A buyer should therefore request current terms from the chosen provider rather than assume a list price represents the final cost.<\/p>\n<h2>What Safety and Licensing Details Need Review?<\/h2>\n<p>Meta says Muse Spark Open received additional safety post-training for open-weight deployment. The company evaluated it under the Meta Superintelligence Scaling Framework and said it met the thresholds required for release. Detailed results are in the model card, while a full evaluation report was still pending in the launch announcement.<\/p>\n<p>The <a href=\"https:\/\/research.meta.ai\/blog\/muse-spark-open\" target=\"_blank\" rel=\"noopener\">official release<\/a> says the detailed evaluation is available in the model card and a full report is coming. The IT lead should read the current model card, usage policy and license together, then document which business uses are permitted.<\/p>\n<h2>What to Do Now<\/h2>\n<ol>\n<li><strong>Calculate the full local deployment commitment.<\/strong> Ask for the purchase price, power and cooling requirements, expected support effort, model-update process and the cost of replacing hardware before approving two Mac Studios.<\/li>\n<li><strong>Assign an owner.<\/strong> Identify the person or service provider responsible for installation, serving, monitoring, backups where applicable, updates and troubleshooting. Confirm that they understand the mixed-hardware and distributed-runtime requirements.<\/li>\n<li><strong>Map where business data goes.<\/strong> For a local setup, establish where weights, prompts, logs and generated files reside, who can access them and how they are removed. For hosted inference, review the provider&#8217;s contract and data-processing terms.<\/li>\n<li><strong>Define output review.<\/strong> Decide which tasks may run without individual review and which require a person to check code changes, customer responses, financial material or other consequential output.<\/li>\n<li><strong>Compare hosted and local costs against real workloads.<\/strong> Include hardware, power, specialist labor and provider usage in the decision. Use current provider terms because public catalogs do not show one consistent first-party price.<\/li>\n<\/ol>\n<h3>Separate Pilot Traffic From Production Work<\/h3>\n<p>Start with a non-production task that does not expose customer records, confidential employee data or sensitive financial information. Suitable examples from the release context are generating tests, summarizing support tickets or scripting report jobs. Measure latency, accuracy, failures and the work required to check each result.<\/p>\n<p>Do not move the workflow into production merely because the pilot completes. First determine whether the output meets a defined quality threshold, whether a person can review it within the available time and whether the team can restore normal operations when the model, runtime or hardware changes.<\/p>\n<h2>FAQ<\/h2>\n<h3>What Is Muse Spark Open?<\/h3>\n<p>Meta describes Muse Spark Open as a natively multimodal open-weight model with 816B total parameters, 42B active parameters and a 1M-token context window, built for coding and agentic work.<\/p>\n<h3>Can Muse Spark Open Run on a Desktop?<\/h3>\n<p>Yes. Meta says Muse Spark Open is the first model in its class to run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework.<\/p>\n<h3>Where Can Developers Download or Access Muse Spark Open?<\/h3>\n<p>Weights are available on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. Hosted inference is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.<\/p>\n<div class=\"source-disclaimer\"><em>This article reports on publicly available information from the sources listed below. We have no reason to believe it is inaccurate, but we have not independently verified every claim, and details may change as the story develops.<\/em><\/div>\n<section class=\"post-sources\">\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/research.meta.ai\/blog\/muse-spark-open\" target=\"_blank\" rel=\"noopener nofollow\">research.meta.ai<\/a><\/li>\n<li><a href=\"https:\/\/benchlm.ai\/models\/muse-spark-open\" target=\"_blank\" rel=\"noopener nofollow\">benchlm.ai<\/a><\/li>\n<li><a href=\"https:\/\/llm-stats.com\/models\/muse-spark-open\" target=\"_blank\" rel=\"noopener nofollow\">llm-stats.com<\/a><\/li>\n<li><a href=\"https:\/\/theairankings.com\/meta\/muse-spark-open\/\" target=\"_blank\" rel=\"noopener nofollow\">theairankings.com<\/a><\/li>\n<li><a href=\"https:\/\/dev.meta.ai\/resources\/blog\/build-with-muse-spark-open\" target=\"_blank\" rel=\"noopener nofollow\">dev.meta.ai<\/a><\/li>\n<\/ul>\n<\/section>\n<p><script type=\"application\/ld+json\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"FAQPage\",\"mainEntity\":[{\"@type\":\"Question\",\"name\":\"What Is Muse Spark Open?\",\"acceptedAnswer\":{\"@type\":\"Answer\",\"text\":\"Meta describes Muse Spark Open as a natively multimodal open-weight model with 816B total parameters, 42B active parameters and a 1M-token context window, built for coding and agentic work.\"}},{\"@type\":\"Question\",\"name\":\"Can Muse Spark Open Run on a Desktop?\",\"acceptedAnswer\":{\"@type\":\"Answer\",\"text\":\"Yes. Meta says Muse Spark Open is the first model in its class to run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework.\"}},{\"@type\":\"Question\",\"name\":\"Where Can Developers Download or Access Muse Spark Open?\",\"acceptedAnswer\":{\"@type\":\"Answer\",\"text\":\"Weights are available on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. Hosted inference is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.\"}}]}]}<\/script><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Meta released Muse Spark Open, an 816B open-weight model it says can run on two M5 Ultra Mac Studios. Small IT teams must weigh setup, data, licensing and<\/p>\n","protected":false},"author":3,"featured_media":761,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-762","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/managedt.com\/blog\/wp-json\/wp\/v2\/posts\/762","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/managedt.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/managedt.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/managedt.com\/blog\/wp-json\/wp\/v2\/comments?post=762"}],"version-history":[{"count":3,"href":"https:\/\/managedt.com\/blog\/wp-json\/wp\/v2\/posts\/762\/revisions"}],"predecessor-version":[{"id":944,"href":"https:\/\/managedt.com\/blog\/wp-json\/wp\/v2\/posts\/762\/revisions\/944"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/managedt.com\/blog\/wp-json\/wp\/v2\/media\/761"}],"wp:attachment":[{"href":"https:\/\/managedt.com\/blog\/wp-json\/wp\/v2\/media?parent=762"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/managedt.com\/blog\/wp-json\/wp\/v2\/categories?post=762"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/managedt.com\/blog\/wp-json\/wp\/v2\/tags?post=762"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}