--- title: "Muse Spark Open Hardware Requirements: GPU, Memory, and the Two-Mac Setup" slug: muse-spark-open-hardware-requirements-gpu-memory description: "Meta's Muse Spark Open is an 816B-parameter open-weight model with weights on Hugging Face. Meta's own memory floors (700 GB at NVFP4, 2 TB at BF16), which GPU nodes clear them, the two-Mac setup, and the license catch." author: "Yotta Labs" date: 2026-10-16 categories: ["Inference"] canonical: https://www.yottalabs.ai/post/muse-spark-open-hardware-requirements-gpu-memory --- # Muse Spark Open Hardware Requirements: GPU, Memory, and the Two-Mac Setup ![](https://cdn.sanity.io/images/wy75wyma/production/bc28e36246864ca070fffb78264525e0395a3c9e-2240x1260.png) *Meta released the weights, the memory figures, and the engine list on the same day. Here's what it takes to run Muse Spark Open, from a single GPU node to the two Mac Studios in Meta's demo.* Meta released Muse Spark Open on October 15, 2026: an 816 billion-parameter Mixture-of-Experts model with 42B active parameters, open weights on Hugging Face, a 1M-token context, and text, image, video, and audio input. Meta calls it the most capable US open-weight model to date. It's the open-weight sibling of the hosted [Muse Spark](https://www.yottalabs.ai/post/meta-muse-spark-multimodal-model-explained-how-it-works-use-cases) line, and it arrives three days after the closed [Muse Wonder](https://www.yottalabs.ai/post/meta-muse-wonder-pricing-api-cost-muse-code-plans) models. Unlike most launches, this one came with the hardware answer attached. The model card states the memory floors directly, names the inference engines, and describes the Mac configuration Meta used to demo it. This post lays those out, maps them onto the GPU nodes people actually rent, and covers the one thing the launch coverage mostly skipped: the license. ## The numbers that matter Two things stand out. Meta published its own memory floors, so the hardware section below is built on Meta's numbers rather than arithmetic. And the license is custom, with conditions that matter if you plan to serve the model to others. More on that below. ## Why 42B active doesn't make it small Muse Spark Open routes each token through 8 of its 256 experts, so about 42B parameters do the work on any given token. That keeps compute per token in the same range as a mid-sized dense model. It does nothing for memory. Every expert has to sit in GPU memory and be ready, because the router picks a different set for each token. Meta's model card says this directly: activating 42B parameters per token does not reduce the requirement to that of a 42B model. It's the same rule that shaped [Kimi K3](https://www.yottalabs.ai/post/kimi-k3-hardware-requirements-gpu-memory-2026) (104B active, 1.56 TB of weights) and [GLM 5.3](https://www.yottalabs.ai/post/glm-5-3-hardware-requirements-gpu-memory-2026) (about 40B active, 744 GB in FP8). Low active parameters save compute, not VRAM. ## Meta's memory floors, and what's behind them The two columns don't match, and the difference is the useful part. At NVFP4, 816B parameters weigh about 408 GB, but Meta says to plan for at least 700 GB. The extra is activations, the perception encoder, and above all KV cache for a 1M-token context, which is expensive on a model with 62 layers. Meta hasn't said what context length its floor assumes, so treat 700 GB as the minimum to load and serve, not the ceiling for long-context work. The practical read: use Meta's floors, not the weight size, when you pick a node. ## Which GPU configurations fit "Fits" means Meta's stated floor fits in total GPU memory. The NVFP4 checkpoint is the one to plan around. It's a single-node model on an 8x H200 node or a 4-GPU Blackwell configuration, and an 8x H100 node doesn't clear Meta's floor. BF16 needs an 8x B300 node at minimum, and realistically a cluster if you want headroom for fine-tuning or long context. That puts Muse Spark Open in the same hardware class as GLM 5.3 and a step below Kimi K3, with the trade that its NVFP4 checkpoint is official rather than a community quantization. ## The two-Mac setup The detail that made the headlines: Meta's demo ran the model on two Mac Studios with M5 Ultra chips, 512 GB of unified memory each, about 1 TB combined, using MLX. That clears the 700 GB NVFP4 floor with room to spare, and it's a rare sight for a model this size. Three things to know before you treat that as a plan. Meta hasn't published token-generation speed for the Mac setup, and unified memory bandwidth is well below an H200 node's, so expect it to be a development and evaluation setup rather than a serving one. The 512 GB Mac Studio configuration wasn't shipping at launch, per the coverage. And two of them plus the networking is a real capital cost against renting a GPU node by the hour, so run the math for your own usage before buying. There is no single-GPU path. The smallest floor is 700 GB, and the largest single GPUs hold under 300 GB. ## Engines and software Meta named the serving stack on day one: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Unsloth for fine-tuning, llama.cpp, and MLX for Apple Silicon. The model card points to SGLang's Docker image for the quickest start, and recommends running at high or xhigh reasoning effort for coding and agent work, with temperature 1.0 and top_p 1.0. For a model this size, SGLang's prefix caching matters: agent loops that resend the same system prompt and tool definitions on every turn get most of that compute back. Our [SGLang guide](https://www.yottalabs.ai/post/what-is-sglang-architecture-performance-and-when-to-use-it) covers how that works and where vLLM still wins, and the [vLLM vs SGLang comparison](https://www.yottalabs.ai/post/vllm-vs-sglang-which-inference-engine-should-you-use-in-2026) covers choosing between them. ## The license Muse Spark Open ships under the Muse Model License, not Apache 2.0 or MIT. You can run it, modify it, fine-tune it, distill it, and redistribute it, subject to the terms. Three of those terms matter for anyone planning a deployment: - **Attribution.** Products built on it must display "Built with Muse Spark Open" in a permitted location. - **Model-as-a-service threshold.** If you offer the model as a service and your affiliates' aggregate revenue exceeds $20 million in any 12-month period, you must notify Meta and sign a separate agreement first. Internal use, research, and products that aren't model serving are exempt. - **Derivatives.** The definition is broad and includes models improved using Spark Open's outputs. For an internal deployment or a product that uses the model rather than resells it, none of that is a blocker. For a hosting business, it is a legal review before launch. Read the license before you build on it. ## What Meta reports on benchmarks All of these are Meta's own numbers, run at Meta's chosen reasoning settings, and the comparison models were run at theirs. Read them as a rough ordering. On Meta's own table the Chinese open flagships stay ahead on coding and on the broad index, and Spark Open leads on the agent and tool-use rows. Combined with native image, video, and audio input, that's the pitch: a US-built open model for multimodal agent work, not a new coding leader. ## What it costs to call instead If you don't want to run it, several hosts picked it up on day one. Together AI's launch pricing was $1.25 per million input tokens and $4.25 per million output, per the launch coverage, which is the same rate as Meta's hosted Muse Spark 1.x. Other hosts named at launch include Fireworks, Groq, Baseten, Modal, Nebius, and OpenRouter, with varying context limits. For comparison, the open flagships on [Yotta AI Gateway](https://www.yottalabs.ai/ai-gateway) run $0.99 in and $2.97 out for DeepSeek V4 Pro and $1.40 in and $4.40 out for GLM 5.3, so Spark Open at Together's rate sits in the same band. Self-hosting an 8x H200 node has to earn its place against those numbers, and the reasons it does are the usual ones: data that can't leave your infrastructure, fine-tuning, and sustained volume. ## Frequently asked questions **How big is Muse Spark Open?** 816 billion total parameters with 42 billion active per token, plus a 1.8B perception encoder, per Meta's model card. Weights are released in BF16 and NVFP4. **How much VRAM does Muse Spark Open need?** Meta states at least 700 GB of accelerator memory for the NVFP4 checkpoint and 2 TB for BF16. Those are Meta's floors, not estimates. **Can I run Muse Spark Open on 8x H100?** Not at Meta's stated floor. Eight H100s hold 640 GB against a 700 GB minimum for NVFP4. **Can I run Muse Spark Open on 8x H200?** Yes, for NVFP4: 1,128 GB of GPU memory against a 700 GB floor, with room for long context. BF16 would not fit. **Can I run Muse Spark Open on a Mac?** Meta demoed it on two M5 Ultra Mac Studios with 512 GB of unified memory each, about 1 TB combined, using MLX. One Mac is not enough at any current configuration, and Meta hasn't published speed figures for the setup. **Is Muse Spark Open open source?** It's open-weight under the Muse Model License, which is a custom license. It allows running, fine-tuning, and redistribution with conditions, including an attribution requirement and a separate agreement for model-as-a-service providers above $20 million in revenue. **Which inference engines support Muse Spark Open?** Meta names SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Unsloth, llama.cpp, and MLX. **Is Muse Spark Open the same as Muse Wonder?** No. Muse Wonder and Muse Wonder Full are closed, hosted models released on October 12. Muse Spark Open is a separate open-weight release. **Is Muse Spark Open multimodal?** It takes text, image, video, and audio as input. Output is text only. **How does Muse Spark Open compare to Mistral Large 4 and Beam on hardware?** It's larger than both: 816B against 1.05T for [Mistral Large 4](https://www.yottalabs.ai/post/mistral-large-4-hardware-requirements-gpu-memory-2026) and 501B for [Reflection Beam](https://www.yottalabs.ai/post/reflection-beam-hardware-requirements-gpu-memory-2026). The difference is that its weights and memory floors are published today, while the other two are still estimates waiting on a release. ## Bottom line Muse Spark Open is an 816 billion-parameter open-weight model with Meta's own memory floors attached: at least 700 GB for the NVFP4 checkpoint, 2 TB for BF16. That makes it an 8x H200 or 4-GPU Blackwell model at 4-bit, an 8x B300 model at full precision, and a two-Mac-Studio model for development if you have about 1 TB of unified memory on your desk. The engines are ready on day one and the weights are public, so the open question isn't hardware. It's the license, which needs a read before any model-as-a-service deployment. If the terms work for you, [8-GPU H200 and B300 nodes by the hour](https://www.yottalabs.ai/pricing) cover both checkpoints, and the open models already on [Yotta AI Gateway](https://www.yottalabs.ai/ai-gateway) give you the price comparison to test it against.