Muse Spark Open can run locally on two Mac Studio M5 Ultra machines, each equipped with 512 GB of unified memory, according to Meta in 2026. But this setup remains difficult for a small or midsize business to reproduce as of October 21, 2026: the required machines have not yet been delivered, their final price is unknown, and no independent review confirms the total cost of operation.
What is Muse Spark Open, and what is this model used for?
Muse Spark Open is an open-weight multimodal artificial intelligence model released by Meta on October 15, 2026. It processes text, images, audio, and video, then generates text. Its announced uses include programming, software tool use, and long-running autonomous tasks.
An open-weight model allows users to download the parameters learned during training and run the model on their own infrastructure. This does not automatically mean “open source”: Muse Spark Open is covered by a specific Meta license, the Muse Model License, rather than a license recognized by the Open Source Initiative.
This distinction changes how a company can integrate, modify, or commercialize the model. Before beginning any development, a lawyer should therefore verify that the intended use complies with Meta’s commercial terms. Downloading the weights does not eliminate contractual risk.
Muse Spark Open has a total of 816 billion parameters, but only 42 billion are activated for each token processed. A token is a small unit of text, often a word fragment. This architecture aims to preserve substantial capacity while limiting the amount of computation used at each step.
How can 816 billion parameters fit on two Mac Studio machines?
Running Muse Spark Open on two Mac Studio machines relies on a sparse mixture of experts architecture, or sparse mixture of experts. Each token uses 8 experts out of 256 in each relevant layer, rather than the entire model. MLX then distributes the weights and computations between the two machines.
The model has 62 transformer layers, a mathematical structure that analyzes relationships among data. The word “expert” does not refer to an independent agent: it is a specialized subnetwork activated by the routing mechanism. The full weights must nevertheless be stored in memory, even though they are not all computed simultaneously.
La mémoire unifiée des puces Apple peut être utilisée par le processeur central et le processeur graphique sans duplication systématique. Meta annonce deux Mac Studio M5 Ultra de 512 Go chacun, soit 1 024 Go cumulés, avec une bande passante de 1,2 To/s par machine. MLX, le framework d’Apple pour le calcul sur puces Apple Silicon, distribue l’inférence entre les ordinateurs grâce à RDMA sur Thunderbolt, un accès direct à la mémoire distante.
Quantization makes this setup possible. It reduces the numerical precision of the weights to decrease their memory footprint, with a potential trade-off in quality or compatibility. Meta provides checkpoints in BF16 and NVIDIA NVFP4; the Mac implementation relies on the MLX stack presented alongside the demonstration.
Can Muse Spark Open really be installed locally in 2026?
Muse Spark Open can technically run locally, but the two-Mac Studio demonstration is not immediately reproducible as of October 21, 2026. Apple plans to make Mac Studio M5 Ultra machines with 512 GB of unified memory available by the end of October 2026, while Meta provides neither throughput for two machines nor the total cost.
Apple reports more than 36 generated tokens per second on a cluster of four Mac Studio machines. This figure should not be mechanically divided by two: speed depends on how the model is partitioned, context length, the network between nodes, and the volume of simultaneous requests. Meta has published no equivalent throughput figure for its two-machine setup.
The context window reaches 1,048,576 tokens, meaning the maximum amount of information the model can consider in a request. This capacity can be used to analyze a large code repository or extensive documents. However, it greatly increases the memory consumed during inference; displaying a maximum capacity therefore does not guarantee that it can be used with good performance in every configuration.
Supported software stacks include MLX, SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, and Unsloth. For a small or midsize business, theoretical compatibility is only the beginning. It is also necessary to plan for monitoring, updates, access management, and a disaster recovery plan suited to the infrastructure.
In the projects we lead, we often see the same blind spot: the team budgets for the hardware but not for operations. MLX expertise, observability, backups, and the time spent diagnosing a drop in throughput can weigh more heavily than the initial installation.
How much does Muse Spark Open cost on Mac, GPU, or API?
The announced minimum cost of two Mac Studio M5 Ultra machines reaches $10,998 before taxes in 2026, excluding the 512 GB option, storage, and networking. Meta’s actual setup cost therefore remains unknown. By contrast, a hosted API charges $5.50 for 1 million input tokens and 1 million output tokens.
Apple sells the Mac Studio M5 Ultra starting at $5,499 before taxes in 2026. This base price does not correspond to the required configuration, since the surcharge for reaching 512 GB of unified memory had not been published as of October 21, 2026. Two optional displays, storage, compatible cables, electricity, maintenance, and administrative time must also be added.
| Option | Configuration or pricing | What this entails |
|---|---|---|
| Two Mac Studio machines | Two M5 Ultra machines with 512 GB; minimum of $10,998 before taxes, excluding the memory option | Feasibility announced by Meta; final price and throughput for two machines not published |
| NVIDIA server in BF16 | At least 2 TB of VRAM according to Meta; Hugging Face test on 32 H200 GPUs | Heavy distributed infrastructure, reserved for equipped teams |
| NVIDIA server in NVFP4 | At least 700 GB of VRAM; eight B200 GPUs in the Hugging Face test | Smaller footprint, but specialized and expensive hardware |
| Hosted API | $1.25 per million input tokens and $4.25 per million output tokens | No server purchase; variable expenses based on usage |
Together AI, Fireworks, and OpenRouter listed the same price in 2026: $1.25 per million input tokens, $0.15 for cached input, and $4.25 per million output tokens. One hundred million input tokens and the same number of output tokens therefore represent $550, excluding ancillary services.
Honestly, an API is often the more rational choice for validating a product or absorbing irregular traffic. Purchasing hardware begins to make sense when volumes are stable, confidentiality requires local execution, and the company already has the necessary expertise. The cost of a potential breach must also be part of the decision, as illustrated by the question of securing data entrusted to a service provider.
When is local deployment preferable to a cloud API?
Un déploiement local de Muse Spark Open se justifie surtout pour des données sensibles, un volume soutenu ou un besoin de maîtrise technique précis. Une API cloud reste préférable pour un prototype, une charge variable ou une équipe sans spécialiste de l’inférence distribuée. Le bon choix dépend du coût total, pas du seul prix par jeton.
Local confidentiality is not automatic. An internal server may still send logs, diagnostic files, or backups to third-party services. It must also be protected against compromised accounts, malicious dependencies, and configuration errors; a simple VPN alone does not guarantee the protection of data used by an AI.
Before choosing, assess the five factors that truly change the budget and risk:
- monthly volume of input, output, and cached tokens;
- la sensibilité des documents au regard du RGPD et des contrats clients ;
- acceptable response time and number of simultaneous users;
- available expertise for operating MLX, vLLM, or GPU infrastructure;
- the terms of the Muse Model License for the intended commercial use.
The obvious choice may be the wrong one. A company handling confidential files might prefer local deployment, but a poorly administered installation may sometimes be less secure than a properly contracted and isolated API. Conversely, routinely sending proprietary code to an external provider without a documented retention policy creates an avoidable risk.
On the agency side, the rule of thumb is to measure a representative workload for a few weeks: request length, outputs, peaks, and latency. This measurement makes it possible to compare the API with the amortized cost of infrastructure, rather than making a decision based on the impressive number of parameters.
What risks need to be addressed before an enterprise deployment?
Professional deployment of Muse Spark Open requires addressing licensing, confidentiality, access security, and service continuity. Open weights cover neither GDPR compliance nor response reliability. Human validation remains necessary for code, automated actions, and sensitive content.
Meta specifically intends the model for tool use and long-running agentic tasks, meaning sequences of actions executed with a certain degree of autonomy. The more a model can modify code, open files, or control software, the more its authorizations must be restricted. An isolated environment, temporary secrets, and approval before any irreversible action limit potential damage.
Meta says it tested Muse Spark Open on complete modifications to the iOS application Ice Cubes and on automations in DaVinci Resolve. These tests had not been independently reproduced in a publication as of October 21, 2026. They demonstrate announced use cases, not a guaranteed level of reliability for your information system.
Governance of generated content deserves the same care: data provenance, validation, retention, and traceability. For published images or videos, watermarking and traceability mechanisms such as SynthID or C2PA address a different question than local hosting, but they can complement the overall framework.
Defining the scope of use, testing with representative data, and calculating the total cost avoids most unpleasant surprises. An outside perspective can be particularly helpful in distinguishing a successful technical demonstration from an architecture that can be operated effectively day to day.
FAQ about Muse Spark Open
Is Muse Spark Open truly open source?
Muse Spark Open is an open-weight model, but its Muse Model License is not an open source license approved by the Open Source Initiative. Meta’s terms must be reviewed before commercial use.
How much memory does Muse Spark Open require?
Meta indicates at least 2 TB of VRAM for the BF16 checkpoint or 700 GB for NVIDIA NVFP4 in 2026. The Mac demonstration relies on two Mac Studio M5 Ultra systems totaling 1,024 GB of unified memory.
Does Muse Spark Open understand images and videos?
Muse Spark Open accepts text, images, audio, and video as input, then produces text. The model can therefore analyze multiple types of content without directly generating an image or video.
What throughput does Muse Spark Open offer on Mac Studio?
Apple reports more than 36 tokens per second on four Mac Studios in 2026. No official throughput for two Mac Studios had been published by Meta as of October 21, 2026.