Meta Muse Spark Open: 816B open weights that run on two Mac Studios

Meta published the Muse Spark Open weights on October 15, 2026

The Meta Muse Spark Open launched on October 15, 2026: an open, natively multimodal model with 816 billion total parameters, 42 billion active per request, and a 1 million-token context window. The announcement positions it as the most capable open-weights model from the United States to date, built for code and agentic work, and it is the first in its class to run on a desktop, using two Mac Studios with the M5 Ultra chip and Apple's MLX framework. The combination of scale and local execution exists because the gap between total and active parameters separates compute cost from the memory requirement, and that separation repositions the choice between a hosted endpoint and your own infrastructure.

Sparse architecture: how 816B processes like 42B

The number that organizes everything in this release is the gap between 816B and 42B. The 816 billion are the model's total parameters, the size of the downloadable artifact. The 42 billion are the active parameters, the fraction that actually processes each token. That is the signature of a mixture-of-experts architecture, and it is what Muse Spark Open 816B exploits: instead of a single dense block, the model keeps a large set of specialized blocks and activates only the ones relevant to each input.

The practical consequence shows up in two places. On compute cost, every generated token pays for the processing of 42B parameters, not 816B, which reduces the arithmetic work per token compared with a dense model of equivalent capability. On memory cost, the math inverts: the full weights have to be available, and that is where the size of the artifact charges its price. Meta's official numbers: the reference weights in BF16 add up to roughly 1.6 TB, and the NVFP4 checkpoint roughly 530 GB, one third of BF16, because only the expert blocks are quantized. The collection on Hugging Face confirms that both formats are available, and the choice between them is exactly the choice between maximum precision and memory feasibility.

Multimodality is native, not an adapter bolted on afterwards: the model processes text, image, video, and audio within the same architecture. For agentic work, that means the reading stages of a long task, a repository with screenshots, meeting recordings, and logs, do not require a separate vision model in front of the language model.

The 1 million-token window completes the agentic package. Context of that order makes it possible to fit entire codebases, or hours of agent interaction, into a single session. The correct reading, however, is the same one we applied to the Gemini 4 Argon 1M window: input capacity is not guaranteed recall. The window says what fits; retrieval quality at positions far from the start of the context is a property that only evaluation can show, and the release includes no independent audit of that point.

On the inference engineering side, Meta shipped the weights alongside custom Helion GPU kernels inside SGLang, with reported speedups of 1.03x to 3.95x at kernel level on NVIDIA and AMD GPUs, and a 3% end-to-end gain on AMD MI350. These are Meta's own numbers, at kernel level, not full-application level. The announcement also lists support for SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon.

The two Mac Studio condition: what is fact and what is context

The most quoted claim of the release deserves to be taken apart. According to Meta, Muse Spark Open is the first model in its class to run on a desktop using two Mac Studios with M5 Ultra and the current version of MLX. Reading the Meta Muse Spark Apple MLX pairing starts with Apple's unified memory: each M5 Ultra makes up to 512 GB of unified memory available to the integrated GPU, at 1.2 TB/s of bandwidth, and two machines add up to 1 TB, enough to load the entire NVFP4-quantized checkpoint (the ~1.6 TB BF16 does not fit).

Three qualifications matter for anyone reading that claim as a purchasing option. First: "two Mac Studios" here means the top configuration, 512 GB per machine; an ordinary Mac Studio does not fit into that equation. Second: the 1.2 TB/s memory bandwidth reads the small fraction each token needs, and that is what makes inference feasible, but generation speed on consumer hardware was not published in the release; anyone who needs throughput will want a measured number before deciding. Third: the demonstrated use cases are Meta's own, running Muse Code on top of Spark Open on the device: real fixes shipped to Ice Cubes, the open source Mastodon client for iOS, with a full build, and end-to-end operation of DaVinci Resolve, with the model reviewing footage frame by frame and cutting vertical crops. These are capability demonstrations, not independent benchmarks.

The correct economic reading is not "a frontier model fits on a desk". It is that the open frontier no longer requires a datacenter and now requires a high-memory dual-workstation setup. The difference between those two sentences is the difference between accessible and merely cheaper than a cluster.

Benchmarks and real-world performance

The official announcement contains no benchmark table and points to a methodology report and to the model card on Hugging Face, with a full evaluation report promised for later. What exists today from primary sources are the agentic demonstrations described above and the positioning statement: most capable open-weights model from the United States as of the date.

Specialized coverage fills part of the gap with third-party numbers. Reporting published during the week, including The Rundown AI and AlphaSignal, cites the Artificial Analysis Intelligence Index with Spark Open at 39 points, ahead of Western open-weights models and behind GLM 5.3 (45), Kimi K3 (44), and Qwen3.8 (40), with the Chinese models keeping their lead on code. That picture is consistent with what the index-based reading of the frontier already showed: the open race continues to be led by China, and Meta's announcement repositions the West in the contest without claiming the lead. Treat those index figures as coverage to be confirmed in the official evaluation report, which has not come out yet.

For engineering teams, the practical implication is expectation calibration: the model is marketed for long-horizon agentic tasks, complex software engineering, and multi-step reasoning, and Meta's demonstrations point in that direction. But no independent code evaluation, such as SWE-bench, accompanied the release. An adoption decision based on a specific benchmark needs to wait for the full report or run its own test on the real workload.

Inference economics: what is known and what was not said

Meta did not disclose API pricing at launch, neither for its own endpoints nor as a reference for the hosted inference partners. Any cost-per-million-tokens table today would be invention. What can be assembled with evidence is the map of consumption options:

The absence of pricing is not a minor detail. The model's cost argument is sparsity: 42B active per token against 816B total means the marginal cost of generation scales with the active fraction, not with the size of the artifact. It is the same economic axis that the cost-per-task analysis already consolidated: active parameters are the best predictor of inference cost, far better than the artifact's marketing number. The release applies that logic at the top of the table: frontier-dense footprint with the cost profile of a much smaller system.

The license is the second economic component of Meta's open model. The weights ship under the Muse Model License, with limited commercial terms, which according to Meta allow free use by many developers. Limited commercial terms mean there are restrictions, and companies with commercial use at scale need to read the license before building on top of it. Open weights are not open source, and that distinction has already decided routing decisions in earlier releases.

Integration and routing in production

The set of consumption paths is the broadest ever seen in an open release from Meta: seven hosted API platforms, four weights repositories, seven inference engines, and four silicon families. For anyone operating a model fleet, each path has a distinct trade-off profile.

The hosted route is the lowest-friction one: create a key with one of the seven providers, point the workload at it, pay per token once pricing is published. The self-hosted route, on GPU, requires memory capacity compatible with the chosen format, roughly 1.6 TB in BF16 or 530 GB in NVFP4, Meta's official numbers, plus vLLM or SGLang operations expertise. The desktop route on two Mac Studio M5 Ultra machines is the hardware novelty, suited to teams that want dedicated inference without a GPU queue, with the caveat that throughput is unpublished.

It is at this point that the decision becomes a matter of gateway and routing, not a binary choice. A real agentic workload mixes profiles: reading and summarization tasks that a small model solves, code engineering tasks that justify a large model, and demand spikes that idle self-hosted infrastructure should not have to absorb on its own. The decision pattern this release reinforces is the same one from when a small model gets close to the leader: capabilities on the table, cost per task measured, and the route chosen by policy, not by habit. The release also does not change the rule of comparing endpoints on the same prompt before deciding; it only adds one more candidate at both ends, the hosted API and self-hosted execution.

Muse Spark Open does not eliminate the opportunity cost of the closed alternative. Meta's on-device demonstrations are real, but the same announcement distributes the model across seven clouds, which indicates that Meta expects high-volume hosted consumption. For most operations, the entry path is the hosted API, with the option of bringing in-house only the workload that justifies it.

Frequently asked questions

Is Muse Spark Open open source?

No. It is an open-weights model under the Muse Model License, with limited commercial terms. The weights can be downloaded and run, and Meta states that use is free for many developers, but the license carries commercial restrictions that need to be read before corporate use at scale.

Do I need two Mac Studios to run the model?

For the announced desktop configuration, yes: two Mac Studios with M5 Ultra and 512 GB of unified memory each, via Apple MLX, adding up to 1 TB. On servers, the BF16 and NVFP4 checkpoints run on engines such as vLLM and SGLang in GPU configurations with compatible memory, roughly 1.6 TB in BF16 or 530 GB in NVFP4, Meta's official numbers, because quantization covers only the experts.

How much does the Muse Spark Open API cost?

Meta did not disclose pricing in the October 15 release. The model is available on seven hosted platforms, Baseten, Fireworks AI, Groq, Modal, Nebius, Together AI, and multi-model API aggregators, and per-million-token pricing on each of them had not been published as of the date of this analysis.

Does the model replace closed frontier models?

No independent data published as of 2026-10-19 supports that reading. Third-party coverage places the model behind GLM 5.3, Kimi K3, and Qwen3.8 on the Artificial Analysis index, and closed frontier models remain ahead on code evaluations. What the release does support is something else: the Western open frontier exists again at the largest scale ever published, with a self-hosted execution path.

Is it worth running locally or using an API?

It depends on the workload. Self-hosted execution makes sense when there is continuous use that pays for the hardware, a data residency requirement, or network isolation. The hosted API wins when demand is variable, when throughput matters, and when the team does not want to operate inference infrastructure. With no published pricing, the economic comparison remains incomplete; measure both routes on your own workload before deciding.

References and further reading

• Meta's official announcement, October 15, 2026
• Development and execution guide, dev.meta.ai
• Weights collection on Hugging Face, including the model card
• PyTorch post about the release
• Evaluation methodology declared by Meta
• Previous analysis: Meta Muse Wonder and LLM routing
• Previous analysis: LiquidAI LFM2.5-24B-A2B and active-parameter efficiency

Next step

Muse Spark Open hands the West an open-weights candidate at the largest scale ever published, with 42B active parameters per token and a self-hosted execution path that no longer requires a datacenter. For the architect operating multiple model routes, the concrete next step is to test the model across its two consumption profiles, hosted endpoint and self-hosted checkpoint, with the same prompts and the same tasks, and measure cost per task before choosing. The Nexforce Router exists precisely for this kind of decision: run the same prompt in parallel across several models, apply routing rules per key, keep a spending cap and failover in place, and let route policy follow measurement, not announcements.