How to Run Muse Spark Open Locally

You can run Muse Spark Open locally, but only if you have roughly 300 gigabytes (GB) or more of combined memory across your system. At Layer3Labs, we build and run AI systems inside other people's businesses, and on a model this size we check memory before choosing an engine or downloading anything. Systems with less memory cannot load this 816-billion-parameter model and must use a hosted Application Programming Interface (API) instead.

Meta released the weights for Muse Spark Open on October 15, 2026. Meta Superintelligence Labs built it. The Meta research blog and the PyTorch engineering update describe a sparse Mixture-of-Experts (MoE) design across 256 experts. The model has 816 billion total parameters, with about 42 billion active per token. All 816 billion parameters must sit in memory even though each token uses about 42 billion, so total memory decides whether your hardware can load it.

Most setups use one of three checkpoints: Meta's full Brain Floating Point 16 (BF16) weights, Meta's 4-bit NVFP4 weights, or a community GGUF build.

Can You Run Muse Spark Open Locally? The Hardware That Fits

Running Muse Spark Open locally requires at least 300 gigabytes (GB) of combined memory across your hardware cluster, which prevents standard single-GPU workstations or personal laptops from loading the weights. Sizing a cluster depends on the precision format you deploy, ranging from heavily quantized community files to full uncompressed checkpoints.

These setups and checkpoints come from the Meta developer blog. The Mac figures are Meta's claim, and cells marked Not published are not stated there.

If your system does not provide at least 300 GB of combined Random Access Memory (RAM) and Video Random Access Memory (VRAM), self-hosting cannot start because the model weights will not fit. Teams with high-memory multi-GPU nodes or networked Apple Silicon machines can run the model on-premise. Smaller teams should use hosted endpoints or see our open-weights models guide for smaller models.

• Memory floor: Requires at least 300 GB combined memory for 2-bit quantization, 530 GB for NVFP4, and 1.6 TB for BF16.
• Mac clustering: Meta claims two 512 GB M5 Ultra Mac Studios run the model via MLX and Thunderbolt 5 connections.
• GPU tiers: NVIDIA Floating Point 4 (NVFP4) fits on four Blackwell Ultra GPUs or eight H200 accelerators.
• Alternative silicon: AMD MI350X/MI355X run FP8 or MXFP4 checkpoints, and Intel offers MXFP4 for four 160 GB cards and INT4 GPTQ for Xeon 6.

Step-by-Step Setup to Run Muse Spark Open Locally

You can set up Muse Spark Open locally in five steps, and the memory check comes first.

Step 1: Check combined system memory against the hardware table. Confirm your machines have enough memory for the checkpoint you plan to use. Quantized GGUF files need about 300 GB of combined RAM and VRAM, NVFP4 needs about 530 GB of VRAM, and BF16 needs about 1.6 TB of memory.

Step 2: Pick a validated inference runtime. Meta lists SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth as compatible engines across NVIDIA, AMD, Intel, and Apple silicon.

Step 3: Download model weights from the official repository. Access the official meta-models/Muse-Spark-Open-816B-A42B Hugging Face repository for BF16 and NVFP4 checkpoints. For quantized GPT-Generated Unified Format (GGUF) builds, use unsloth/Muse-Spark-Open-816B-A42B-GGUF or ggml-org/Muse-Spark-Open-816B-A42B-GGUF.

Step 4: Start the inference server with exact parallelism parameters. For SGLang deployments, pass the flags --tp-size 4 --ep-size 4 to distribute tensor and expert parallelism across four accelerators. For vLLM, specify tensor-parallel size 4 and set the maximum model length to 1048576 tokens. For TensorRT-LLM on a four-GPU Blackwell node, configure the engine to distribute the 256 experts across the GPUs.

Step 5: Connect your developer tools to the local endpoint. Point coding agents, orchestration frameworks, or OpenAI-compatible client libraries to your server's local IP address and port to handle requests offline.

Running Muse Spark Open on Mac Studio With MLX

Clustering Apple Mac Studio hardware provides a non-server path to run Muse Spark Open locally. Meta states that the model runs across two Apple M5 Ultra Mac Studios with 512 GB of unified memory each. Meta also lists four units with 256 GB each, and both layouts give 1 TB of total memory.

The hardware configuration pairs the NVFP4 checkpoint with Apple MLX, linking the physical machines through Thunderbolt 5 cables operating with Remote Direct Memory Access (RDMA). Meta demonstrated on-device execution of Muse Code using this specific multi-machine layout. Readers can explore the architectural capabilities of that coding assistant in our Muse Code guide.

Unified memory makes this desktop setup viable. Unlike traditional PC architectures where the Central Processing Unit (CPU) and Graphics Processing Unit (GPU) maintain separate memory banks over PCIe buses, Apple Silicon shares a single high-bandwidth memory pool. This shared design allows two M5 Ultra desktop units to hold the 530 GB parameter footprint. However, all Mac performance claims originate solely from Meta's launch demonstration, as independent generation speed benchmarks do not exist yet.

• Cluster size: Two 512 GB M5 Ultra units or four 256 GB units give a 1 TB unified memory pool.
• Interconnect: Machines connect over Thunderbolt 5 with RDMA.
• Software framework: Uses Apple MLX to serve the experts-only NVFP4 checkpoint locally.
• Benchmark status: Real-world tokens-per-second numbers remain unverified by independent testers.

Running Muse Spark Open on GPU Servers

Enterprise deployments of Muse Spark Open rely on distributed GPU server clusters managed by vLLM, SGLang, or TensorRT-LLM. The 816-billion-parameter MoE architecture requires splitting both model layers and individual expert weights across multiple accelerator cards to sustain high token throughput. Teams planning production clusters can compare these infrastructure demands with our guides on how to run Kimi K3 locally and how to run DeepSeek locally.

On NVIDIA hardware, the NVFP4 checkpoint applies experts-only quantization to compress the memory footprint to roughly 530 GB. The NVFP4 checkpoint fits on four Blackwell Ultra GPUs or eight H200 GPUs. The unquantized BF16 checkpoint is roughly 1.6 TB, and Meta does not publish a GPU layout for it. Meta's example SGLang settings are --tp-size 4 --ep-size 4, and its TensorRT-LLM example spreads the 256 experts across a single four-GPU Blackwell node.

On AMD hardware, Meta lists an 8-bit Floating Point (FP8) checkpoint for Instinct MI355X with vLLM and a Microscaling 4-bit Floating Point (MXFP4) checkpoint for MI350X and MI355X. Intel MXFP4 targets clusters of four 160 GB accelerators, and 4-bit Integer (INT4) GPTQ enables offline inference on Intel Xeon 6 processors.

Hardware acceleration kernels deliver measurable speed gains in SGLang. Helion GPU kernels integrated into SGLang produce 1.03x to 3.95x kernel-level speedups on NVIDIA and AMD silicon, resulting in a 3 percent end-to-end throughput increase on AMD MI350 hardware. Additionally, DFlash V2 speculative decoding in SGLang generates 3.7x single-request throughput gains on AMD MI355X accelerators. General cluster networking and storage practices are covered in our guide to running open-weights models.

• NVIDIA allocations: Four Blackwell Ultra GPUs or eight H200 GPUs hold the ~530 GB NVFP4 checkpoint.
• AMD support: Instinct MI350X and MI355X run MXFP4 and FP8 checkpoints using vLLM and SGLang.
• Kernel speedups: Helion kernels in SGLang deliver up to 3.95x kernel gains and 3% end-to-end improvements on AMD MI350.
• Speculative decoding: DFlash V2 in SGLang provides 3.7x single-request throughput on AMD MI355X hardware.

Support Status for Muse Spark Open in Ollama

Meta does not provide an official Ollama model image or modelfile for Muse Spark Open.

The closest practical path to local execution without enterprise server runtimes is running community GGUF checkpoints with llama.cpp or Unsloth. Developers can inspect our Ollama setup guide and llama.cpp explainer to understand the architectural differences between these runtime tools.

Unsloth publishes community GGUF builds quantized down to approximately 2-bit precision. These files reduce the required memory to roughly 300 GB of combined RAM and VRAM, enabling execution on high-memory workstation hardware containing 256 GB of host memory supplemented by graphics memory.

When a Hosted API Makes More Sense

Hosted Application Programming Interface (API) endpoints offer a practical alternative when local hardware budgets cannot accommodate hundreds of gigabytes of memory. Querying cloud endpoints bypasses GPU procurement lead times, node configuration, electricity demands, and cluster maintenance overhead. Meta named seven hosted API partners at launch: Baseten, Fireworks AI, Groq (available in limited preview), Modal, Nebius, OpenRouter, and Together AI.

Meta published no standard per-token pricing table for Muse Spark Open. Each cloud provider sets independent token rates.

Hosted infrastructure is the logical path for organizations that lack dedicated machine learning operations staff or multi-GPU servers. Operating an 816B model in-house requires continuous system administration and hardware monitoring. If your organization handles workloads without rigid air-gapped data residency mandates, hosted endpoints let you use the model without buying hardware. To explore Meta's closed commercial variants, review our Muse Spark 1.3 explainer.

• Official cloud providers: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
• Variable pricing: Meta does not establish standard pricing, leaving token rates to individual providers.
• Operational savings: Hosted endpoints eliminate multi-GPU power, cooling, and cluster maintenance costs.
• Target workloads: Ideal for development prototyping, variable traffic volumes, and teams without dedicated cluster engineers.

Fine-Tuning Muse Spark Open After Deployment

Fine-tuning Muse Spark Open requires specialized distributed training software capable of sharding model states across high-memory GPU clusters. Meta names three training tools in its launch material. TorchTitan covers Supervised Fine-Tuning (SFT), Low-Rank Adaptation (LoRA) and Reinforcement Learning (RL), and NVIDIA NeMo AutoModel covers full SFT and LoRA. Hugging Face SFTTrainer is the third option. Cloud platforms Fireworks AI and Together AI also offer hosted LoRA training environments. Our open-weights fine-tuning guide explains parameter-efficient training mechanics in detail.

Customizing Muse Spark Open binds teams to the Muse Model License. Meta grants royalty-free rights to use, copy, modify, fine-tune, distill, and distribute derivative models, but users must comply with explicit commercial and branding rules. You must prominently display the phrase "Built with Muse Spark Open" in your product interface, documentation, website, or settings menu. The license does not meet the Open Source Initiative (OSI) definition of open source, so "open weights" is the accurate term.

The license imposes a revenue condition focused on Model-as-a-Service (MaaS) offerings. If an organization sells third-party API access to model outputs and exceeds US$20 million in aggregate gross revenue over any consecutive 12-month period, it must notify Meta and sign a separate commercial agreement. Internal use, research, and your own products that run the model without selling API access to it fall outside this revenue condition. License terms are examined thoroughly in our Muse Spark Open weights guide.

Deployments must also adhere to Meta's Acceptable Use Policy. Prohibited activities include malware creation, weapons development, critical infrastructure attacks, fraud, disinformation campaigns, and misrepresenting synthetic text as human-generated.

• Training toolkits: TorchTitan (SFT, LoRA, RL), NVIDIA NeMo AutoModel, and Hugging Face SFTTrainer.
• Attribution duty: Display "Built with Muse Spark Open" in the product interface, documentation, website, or settings menu.
• Commercial limit: Model-as-a-Service businesses generating over US$20 million gross revenue require a separate Meta agreement.
• Acceptable use: Prohibits military uses, malware development, infrastructure attacks, and impersonation.

What you need to run Muse Spark Open yourself

Muse Spark Open is a frontier-scale Mixture-of-Experts model, so "running it yourself" is a real infrastructure decision — not something a single laptop or gaming GPU can do. Match the path below to how seriously you need to self-host. For most teams the API or rented GPUs are the right answer; buying hardware only pays off at steady, high volume or when your data can never leave your walls.

Once Muse Spark Open is running, the fastest way to put it to work day to day is inside Cursor — point it at the model through OpenRouter as a custom model. And if you would rather run a model on one affordable box, see Best mini PCs for local AI and Local AI hardware calculator.

NVIDIA RTX 6000 Ada (48GB)

Power users and small clusters that want cards they own

View on Amazon →