--- title: Muse Spark Open runs on Mac Studio with MLX and RDMA... description: Muse Spark Open, an 816B-parameter mixture-of-experts model, now runs locally across clustered Mac Studios using MLX v0.30.1+ and a new RDMA-over-Thunderbolt... canonical: https://daily.dev/posts/muse-spark-open-runs-on-mac-studio-with-mlx-and-rdma-over-thunderbolt-tgbqazw0h twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Muse Spark Open runs on Mac Studio with MLX and RDMA over Thunderbolt | daily.dev og:description: Muse Spark Open, an 816B-parameter mixture-of-experts model, now runs locally across clustered Mac Studios using MLX v0.30.1+ and a new RDMA-over-Thunderbolt... og:url: https://daily.dev/posts/muse-spark-open-runs-on-mac-studio-with-mlx-and-rdma-over-thunderbolt-tgbqazw0h og:image: https://api.daily.dev/og/posts/TGbqaZW0H.png og:image:alt: Muse Spark Open runs on Mac Studio with MLX and RDMA over Thunderbolt og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Muse Spark Open runs on Mac Studio with MLX and RDMA over Thunderbolt **[Collections](https://daily.dev/sources/collections)** · 4 min read · 0 upvotes · 0 comments ## Summary Muse Spark Open, an 816B-parameter mixture-of-experts model, now runs locally across clustered Mac Studios using MLX v0.30.1+ and a new RDMA-over-Thunderbolt driver in macOS 26.2. A four-machine cluster generates over 36 tokens per second. The new M5 Ultra chip delivers up to 4.6x faster time-to-first-token and 1.6x faster generation than M3 Ultra, with 1.2 TB/s memory bandwidth. MLX's jaccl collectives enable pipeline and tensor parallelism across multiple Macs since a single Studio tops out at 512GB RAM. Apple's write-up includes mlx_lm.benchmark and mlx.launch command examples and shows serving a model locally with the OpenCode agentic coding tool. ## Content Meta has released the weights for Muse Spark Open, a multimodal mixture-of-experts model it calls the most capable US open-weight model to date. It has 816B total parameters with 42B active per token, and a 1M-token context window. It takes text, images, video and audio as input and outputs text. Meta built it for coding and long-horizon agentic work such as multi-step planning, tool use and recovering from failures. ## Architecture and relationship to Meta's closed models Each of the 62 layers holds 256 experts and routes every token through 8. You need the memory of an 816B model, but it generates at roughly the speed of a 42B one. The open version shares a base model with Muse Spark 1.3 and got extra safety post-training so it holds up as a standalone model without system guardrails. It was evaluated under Meta's Superintelligence Scaling Framework. The release is a partial reversal of the closed-frontier approach Meta took in April. It is still not the frontier. Meta describes the model as weaker than its closed Muse Wonder, which launched days earlier with no published benchmarks. Alexandr Wang said Meta will generally not open-source its most capable models, citing safety and capability concerns. That sits oddly next to Zuckerberg's earlier remark that it is dangerous for labs to withhold their most advanced models. This is the second open release in the Muse line, after Glimmer in August. ## License The weights come under the new Muse Model License. Companies above $20M in aggregate gross revenue need a separate agreement with Meta for Model-as-a-Service use, and the license requires a "Built with Muse Spark Open" attribution. So it is open-weight, not open-license, much like Moonshot's Kimi K3 license. ## Benchmarks Results are mixed. - Meta puts it at 39 on the Artificial Analysis Intelligence Index. That leads US rivals but trails China's GLM 5.3 (45) and Kimi K3 (44). - It beats the Chinese models on agentic and tool-use tests (JobBench, MCP-Atlas) and roughly ties on SciCode. - It trails on coding. Terminal-Bench 2.1 is 76.4 against 85.0 and 83.9, and DeepSWE 1.1 is 60 against 68, a gap that hasn't moved since April. ## Where to get it - **Hosted APIs:** Baseten, Fireworks AI, Groq, Modal (shared endpoint, per-token billing), Nebius, OpenRouter and Together AI. - **Weights:** Hugging Face, Amazon SageMaker JumpStart (BF16 and NVFP4 variants), Google Cloud Model Garden and Microsoft Foundry. - **Inference engines:** vLLM had day-0 support, verified on NVIDIA and AMD GPUs and TPUs, with BF16, NVFP4 and MXFP4 and speculative decoding on CUDA and ROCm. SGLang, TensorRT-LLM, llama.cpp and transformers also support it. - **Enterprise:** Dell lists it on its enterprise platform. PyTorch integrated Helion, its kernel-authoring DSL, into SGLang for 1.03-3.95x kernel speedups on NVIDIA GB300 and AMD MI350, and about 3% end-to-end throughput. TorchTitan supports fine-tuning, LoRA and reinforcement learning out of the box. On GB300 NVL72, NVIDIA reports up to 50K tokens per second per GPU and 600 tokens per second per user. It also ships a NIM container and post-training tools through NeMo AutoModel and Megatron Bridge. ## Running it yourself This is not a laptop model. You need at least 700 GB of GPU memory in NVFP4 or about 2 TB in BF16. Meta's pitch is the desktop: two Apple M5 Ultra Mac Studios with 512 GB of unified memory each, linked with MLX. The M5 Ultra has 1.2 TB/s of memory bandwidth, and Apple's RDMA-over-Thunderbolt driver (macOS 26.2) lets the machines share the model. Apple reports over 36 tokens per second on a cluster of four. Mark Zuckerberg said the model runs well on the new Mac Studios thanks to work with @johnternus and Apple. Community quants appeared quickly. A 3-bit quant took about 344 GB and ran at about 21 tokens per second on a 512GB M3 Ultra. Unsloth's dynamic quants start at 176GB for UD-IQ1_S, and its 2-bit GGUFs run on 256GB of RAM. ggml-org published a Q2_K GGUF that people are serving across two Mac Studios with llama serve. Meta also paired the release with Muse Code, its AI coding terminal. Demos included fixing a Mastodon client and operating DaVinci Resolve. ## Reaction Early partners Factory and Ollama praised its tool-calling accuracy. Aaron Levie called it a big step for US open source, since enterprises can use frontier models for orchestration and cheaper tuned models for repeatable work. Polymarket put Meta's odds of leading the AI race at year's end at 3%. I think the interesting question is whether a model that isn't the best open model, and isn't Meta's best model, still matters. For people who want agents running privately on their own hardware, probably yes, even if the hardware costs as much as a car. ## Questions this post answers ### How fast can a cluster of Mac Studios generate tokens for an 816B parameter mixture-of-experts model? A cluster of four Mac Studios generates more than 36 tokens per second running Muse Spark Open, an 816B-parameter mixture-of-experts model. This is made possible by MLX v0.30.1 and later, a low-latency RDMA-over-Thunderbolt driver available in macOS 26.2, and MLX's jaccl communication collectives supporting pipeline and tensor parallelism across machines. _daily.dev gathers hardware benchmarks like this for developers weighing local inference clusters against cloud GPUs._ ### Why do you need multiple Mac Studios to run a very large language model locally? A single Mac Studio tops out at 512GB of RAM, which isn't enough to hold an 816B-parameter mixture-of-experts model like Muse Spark Open. Clustering multiple Mac Studios works thanks to a low-latency RDMA-over-Thunderbolt driver in macOS 26.2 and MLX's jaccl collectives, which support pipeline and tensor parallelism across several machines. _Developers scaling local model inference track memory and clustering constraints like these on daily.dev._ ### How much faster is the M5 Ultra chip than the M3 Ultra for running large language models? The M5 Ultra's GPU Neural Accelerators deliver up to 4.6x faster time-to-first-token and 1.6x faster generation compared to the M3 Ultra. Its memory bandwidth reaches 1.2 TB/s, which is the main factor helping token generation speed when running large models locally. _Engineers choosing hardware for local AI workloads compare chip generation gains like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 297 comments across hackernews, x (as of 2026-10-15). **TL;DR:** Reaction centers on the hardware spectacle (dual Mac Studios, Thunderbolt RDMA) and the huge 1M-token context, with excitement about open weights tempered by notes that it still trails Chinese rivals on benchmarks and that the license isn't fully open. Muse Spark Open: Discussion is dominated by skepticism toward Meta's framing of Muse as 'sticky' and the claim it will win consumer AI, alongside pointed critiques of the license's openness and the model's underwhelming benchmark performance against rivals like GLM 5.3 Flash. **Sentiment:** 24% positive · 35% mixed · 41% skeptical **The case for** - The long 1M-token context is seen as genuinely useful for fitting whole codebases into a single prompt. - Some argue personal AI agents in general (including Muse) are inherently sticky because exported memory/context is hard to transfer elsewhere. - Many see it as a strong flex for US open-weight AI and a notable release to track. - A few note the license is fairly permissive, allowing commercial use short of reselling the model as an API service above a $20M revenue threshold. - Free weights for smaller companies are appreciated even if larger ones need a separate deal with Meta. **The pushback** - It still trails other open models by a noticeable margin on the Artificial Analysis Intelligence Index despite being the best in the US. - Many dispute that Muse is actually 'sticky', pointing out it lacks network effects compared to true social platforms like Instagram or WhatsApp. - Some question whether giving away models is just a sales funnel for more hardware purchases. - Benchmarks cited (Artificial Analysis) reportedly show it underperforming open-weight competitors like GLM 5.3 Flash and Qwen3.8 Flash Next. - Prediction markets are cited as giving Meta only a small chance of having the best model by end of 2026, suggesting skepticism about its competitive position. **By community** - hackernews (heated): Heavy disagreement over whether Meta's Muse ecosystem is genuinely 'sticky' or whether its consumer AI dominance claim is overstated, plus a parallel debate over licensing openness and underwhelming benchmarks. - x (mixed): Excitement about the scale and context window mixes with practical cost questions about replicating the Mac Studio setup and skepticism about Meta's actual standing versus rivals. **Hottest debate:** Muse Spark Open: Whether Muse's download numbers and agent-memory lock-in actually constitute meaningful 'stickiness' or whether it's just hype that will fade without real network effects. **Open questions** - How much will the dual Mac Studio setup actually cost once fully configured and available? - Does Muse actually retain users long-term, or will downloads taper off once novelty wears off? - How does it perform in real agentic tool-calling tasks rather than just benchmark scores? - How does Muse Spark Open really compare in production use across different reasoning/effort levels, not just top-tier benchmarks? - Will the 'Model as a Service' licensing clause meaningfully limit adoption by larger commercial users? **Highlights** > @TheRundownAI nvidia giving away models so you buy more nvidia chips. the freebie has a sales funnel on the end of it > — [rayvahey on x · 1 points](https://x.com/rayvahey/status/2110790397978976684) > @Polymarket Free weights for startups, but anyone over $20M in revenue needs a separate deal with Meta. So the license only gates companies big enough to compete with Meta. Still glad to see an 816B model on Hugging Face. > — [DoDataThings on x](https://x.com/DoDataThings/status/2110801297482084384) > @Polymarket Open weights, long odds. Polymarket gives Meta about a 3% chance of having the best model at the end of 2026. Anthropic and Google split most of the rest > — [polyrayxyz on x](https://x.com/polyrayxyz/status/2110793993411830182) > @TheRundownAI For agent work the index score matters less than tool-calling reliability. Quick test before switching: run one job you already do every week (pull a CSV, summarize it, draft the email) 5 times on the new model and count how many finish without a retry or a fix. > — [hirethebot on x](https://x.com/hirethebot/status/2110821526043775405) > @afrikha44 @Polymarket Two M5 Ultra Mac Studios with 512 GB each: 512 GB configs ship late this month, prices not fully listed yet. 256 GB high-end ones run ~$11k–$18k, so expect $20k+ each (~$40k+ for the pair). NVIDIA BF16 (8× B300): roughly $300k–$650k for a full system. NVFP4 4-GPU Blackwell node: > — [grok on x](https://x.com/grok/status/2110815909426102682) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50107379) · 373 points · 260 comments - [x](https://x.com/Polymarket/status/2110791077699469713) · 264 points · 29 comments - [x](https://x.com/TheRundownAI/status/2110789794905174243) · 10 points · 5 comments - [hackernews](https://news.ycombinator.com/item?id=50106620) · 70 points · 2 comments - [x](https://x.com/UnslothAI/status/2110787461270171981) · 30 points · 1 comments --- Tags: [#local-ai](https://daily.dev/tags/local-ai), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts), [#mlx](https://daily.dev/tags/mlx) [View this post on daily.dev](https://daily.dev/posts/muse-spark-open-runs-on-mac-studio-with-mlx-and-rdma-over-thunderbolt-tgbqazw0h) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Muse Spark Open runs on Mac Studio with MLX and RDMA over Thunderbolt","url":"https://daily.dev/posts/muse-spark-open-runs-on-mac-studio-with-mlx-and-rdma-over-thunderbolt-tgbqazw0h","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/muse-spark-open-runs-on-mac-studio-with-mlx-and-rdma-over-thunderbolt-tgbqazw0h"},"datePublished":"2026-10-15T17:14:16.104Z","dateModified":"2026-10-15T20:59:37.877Z","description":"Muse Spark Open, an 816B-parameter mixture-of-experts model, now runs locally across clustered Mac Studios using MLX v0.30.1+ and a new RDMA-over-Thunderbolt...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/f048f574544c2abecb3593667fa68293?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/f048f574544c2abecb3593667fa68293?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/muse-spark-open-runs-on-mac-studio-with-mlx-and-rdma-over-thunderbolt-tgbqazw0h","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"local-ai,mixture-of-experts,mlx","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Muse Spark Open runs on Mac Studio with MLX and RDMA over Thunderbolt"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/muse-spark-open-runs-on-mac-studio-with-mlx-and-rdma-over-thunderbolt-tgbqazw0h#faq","mainEntity":[{"@type":"Question","name":"How fast can a cluster of Mac Studios generate tokens for an 816B parameter mixture-of-experts model?","acceptedAnswer":{"@type":"Answer","text":"A cluster of four Mac Studios generates more than 36 tokens per second running Muse Spark Open, an 816B-parameter mixture-of-experts model. This is made possible by MLX v0.30.1 and later, a low-latency RDMA-over-Thunderbolt driver available in macOS 26.2, and MLX's jaccl communication collectives supporting pipeline and tensor parallelism across machines. daily.dev gathers hardware benchmarks like this for developers weighing local inference clusters against cloud GPUs."}},{"@type":"Question","name":"Why do you need multiple Mac Studios to run a very large language model locally?","acceptedAnswer":{"@type":"Answer","text":"A single Mac Studio tops out at 512GB of RAM, which isn't enough to hold an 816B-parameter mixture-of-experts model like Muse Spark Open. Clustering multiple Mac Studios works thanks to a low-latency RDMA-over-Thunderbolt driver in macOS 26.2 and MLX's jaccl collectives, which support pipeline and tensor parallelism across several machines. Developers scaling local model inference track memory and clustering constraints like these on daily.dev."}},{"@type":"Question","name":"How much faster is the M5 Ultra chip than the M3 Ultra for running large language models?","acceptedAnswer":{"@type":"Answer","text":"The M5 Ultra's GPU Neural Accelerators deliver up to 4.6x faster time-to-first-token and 1.6x faster generation compared to the M3 Ultra. Its memory bandwidth reaches 1.2 TB/s, which is the main factor helping token generation speed when running large models locally. Engineers choosing hardware for local AI workloads compare chip generation gains like this on daily.dev."}}]} ```