Meta Put An 816B Open-Weight Model On Two Mac Studios: What Muse Spark Open Actually Delivers

Posted on by
Contents show
A flat vector illustration of two identical silver computer towers connected by a thin glowing teal line with a warm amber spark floating between them, representing Muse Spark Open, an 816B parameter open-weight model, running across two Mac Studios

Meta released Muse Spark Open on 15 October 2026: an open-weight, natively multimodal mixture-of-experts model with 816 billion total parameters, about 42 billion active per token, a 1M-token context window, and the claim of being the most capable US open-weight model to date. The claim is doing real work in that sentence, because the open-weight tier has been dominated by Chinese labs through 2026, and Muse Spark Open is the first US checkpoint in months that can sit at that table. The headline that travelled was the hardware story, Muse Spark Open is optimized to run locally on two Mac Studios with M5 Ultra chips, a terabyte of unified memory behaving like one machine. The substance that matters more is in the model card, and it is refreshingly double-edged: Meta's own tables show the model leading its open-weight tier on agentic tool use and safety, while trailing the strongest Chinese open models on coding, the very use case it was built for.

This article works from the Hugging Face model card, the license text and the architecture details, all read directly, and separates what Muse Spark Open delivers from what the announcement implies. It follows this site's earlier Muse coverage, the closed Muse Wonder, and extends the open-weights thread that has run through the autumn releases.

What is actually in the checkpoint

The architecture of Muse Spark Open is a sparse mixture-of-experts transformer with a perception encoder in front. The language model has 62 layers, a hidden size of 8192, grouped-query attention with 128 query heads and 16 key-value heads, and each layer replaces a dense feed-forward block with an MoE block of 256 experts, of which a router selects 8 per token. The experts hold roughly 97 percent of the parameters, which is what lets the model carry 816 billion parameters while activating about 42 billion per token.

Multimodality in Muse Spark Open is native rather than bolted on. Images and video frames pass through a 1.8B-parameter perception encoder; audio is tokenized at one token every 80 milliseconds. Output is text. Context is 1 million tokens at inference, with a vocabulary of 202,048, and Meta reports training data drawn from more than 100 languages. Reasoning cannot be switched off, but effort is adjustable across four levels, low, medium, high and xhigh, with xhigh the default and recommended setting for coding and agentic work.

Two Muse Spark Open checkpoints ship: BF16, which the model card says requires at least 2TB of VRAM or unified memory, and NVFP4, which fits a single B200 node at around 700GB. The KV cache is comparatively small for a model of this size, about 250KB per token in BF16, or roughly 266GB for the full 1M-token context, which is what makes the million-token window operationally realistic rather than theoretical. For local deployment, the model is optimized for two Mac Studios with M5 Ultra using MLX, each contributing up to 512GB of unified memory. Quantization lowers the floor further: a 3-bit quant of Muse Spark Open reportedly runs in about 344GB on a single 512GB machine. Supported serving frameworks include SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth, with API access listed through seven providers and downloads through four clouds.

The benchmarks: leading where it aims first, trailing where it aims second

All published numbers are Meta's own, run at xhigh effort, and the model card says independent replications do not yet exist. Within that caveat, the pattern is consistent and honest.

Agentic tool use is the strongest area. Muse Spark Open scores 89.2 on MCP-Atlas, ahead of every model in Meta's comparison table, and tops JobBench at 61.8, narrowly ahead of GLM 5.3's 61.4. On HLE with tools it reaches 61.0, and on ZeroBench with tools it posts 50.0 against 41.0 for the nearest competitor. If the claim is that this is the most agentic open-weight model available, the tables support it.

Coding trails the Chinese leaders. On Terminal-Bench 2.1 the model scores 76.4, against 85.0 for Kimi K3 and 83.9 for GLM 5.3. On DeepSWE 1.1 it posts 60 against 68 for Kimi K3. On GPQA Diamond it reaches 89.0 against 93.5 for two competitors. The overall Artificial Analysis Intelligence Index score is 39, against 45 for GLM 5.3 and 44 for Kimi K3. Meta's card does not hide these numbers, and the gap on coding, the stated training focus, is the single most important fact for any team choosing between this model and the alternatives.

Safety is the standout. On StrongREJECT v2 adaptive attacks, where lower is better, Muse Spark Open scores 0.5 against 62.9 to 86.4 for the other models listed. On cyber misuse benchmarks queried the way a developer would query an API model with no system-level safeguards, the model refused 99.2 percent of jailbreaks on Meta's internal benchmark and 94.8 percent aggregated across public sets, where the card reports Kimi K3 at 75.5 and 70.2 percent. The open-weight version received additional safety post-training specifically for running standalone without system guardrails, and those numbers are the evidence for it.

The honest summary is that Muse Spark Open is simultaneously the leading open-weight model on agentic tool use and safety, and a few points behind the tier's best on coding and general reasoning. Whether that trade is right for a given team depends on the workload, and Muse Spark Open's trade is stated in Meta's own data rather than in criticism, which is the correct way for a model card to argue.

The two-Mac story, read carefully

The desktop claim deserves precision, because it is both real and easy to overstate. What Meta demonstrated is Muse Spark Open running locally on two Mac Studios with M5 Ultra chips, each contributing up to 512GB of unified memory, connected so that they behave like one machine with a terabyte of memory. That is a genuine local path for frontier-class weights, and it is the first time this size of model has been optimized for Apple silicon in that way.

It is also a two-machine setup costing roughly the price of a car, not a workstation upgrade, and the Muse Spark Open footprint rules out every smaller local rig. Coverage of the launch made the same point from the other direction: the memory floor is about a terabyte, four 256GB machines also work at a substantially higher cost, and discussion in local-model communities has focused on the quantized paths that bring the floor down to a single 512GB machine. For most teams, the practical deployment is hosted, via one of the listed API providers, or a GPU node running the NVFP4 checkpoint. The local story matters as a proof of where unified memory is taking personal compute, and as an option for organisations whose token spending justifies Muse Spark Open hardware, not as a hobbyist path.

The economics comparison is worth making against the site's running coverage of inference costs. A team burning through API tokens on agentic workloads has two lever families: route smarter, as in the model routing guide, or own the capacity. Muse Spark Open is the most concrete data point yet in the own-the-capacity direction for teams inside the Apple ecosystem, with the caveat that the model card's own numbers say the coding workload may still belong to a competitor's weights. There is also a middle path the family enables: serve the coding traffic on whichever open weights win the benchmarks, and use Muse Spark Open for the agentic and multimodal work where it leads, with the shared tokenizer keeping routing between them cheap.

The license: open weights, conditional terms

The Muse Model License that governs Muse Spark Open is permissive in form and conditional in substance. It grants use, modification, fine-tuning, distillation, distribution and sale. It adds three conditions that anyone adopting the model should read before building.

First, attribution: products using the model or any derivative must prominently display “Built with Muse Spark Open,” and derivatives are defined broadly, including models trained or distilled on the model's outputs. That last clause reaches further than many builders will expect: a fine-tune trained on Spark-generated data is a derivative.

Second, the MaaS threshold: a company offering model access through an API, whose aggregate gross revenue exceeds 20 million dollars in any twelve-month period, must notify Meta and sign a separate agreement before offering the model or derivatives as part of that business. Internal use, research and non-API products are exempt. This is the Llama-style revenue trigger, softened for most companies but real at scale.

Third, the usage policy, which excludes military applications, users under 18, and the usual categories of harm. Access on Hugging Face is gated behind a form collecting identity details. The terms are less restrictive than past Llama licenses for most developers, but they are not Apache 2.0 or MIT, and teams that need unencumbered weights have alternatives in this tier. Anyone adopting Muse Spark Open should price in the attribution requirement as a design constraint, because it applies to user interfaces, documentation and settings menus, not only to About pages.

The family context and what it means for positioning

Muse Spark Open does not arrive alone, and the family structure clarifies what it is for. It is the full-size member of the Muse series, sharing a tokenizer and vocabulary with Muse Glimmer, the small dense model released in August, and sitting below Muse Wonder, Meta's most capable model, which remains closed-weight and API-only. The naming is the strategy: Wonder holds the frontier, Spark Open seeds the ecosystem, and Glimmer gives the same tokenizer a cheap local anchor.

That arrangement has practical consequences for builders. A team can prototype against Wonder's quality, deploy against Spark Open's weights, and run lightweight routing or classification against Glimmer locally, all within one token space and one tool ecosystem. Distillation down the family is explicitly licensed, which makes the open-weight checkpoint usable as a data engine for smaller custom models, and the model card lists synthetic data generation and LLM-as-a-judge work among its intended uses. Those are the jobs a 42B-active sparse model can hold continuously, where its 816B siblings' quality is preserved through the training pipeline rather than through serving.

The positioning also explains the safety investment. A closed model sits behind a provider's guardrails, so its post-training can assume a system around it. An open-weight checkpoint gets deployed by whoever downloads it, in whatever scaffold they build, so the model itself must carry more of the safety load. The additional post-training for standalone use, and the published adaptive-attack numbers, are the direct response to that difference, and they are as much about protecting the ecosystem's smallest deployers as about protecting Meta's reputation.

For the competitive read, the family framing softens the benchmark gaps without erasing them. Muse Spark Open does not need to beat Kimi K3 at Terminal-Bench to justify its existence, because its jobs are agency, multimodality, ecosystem and distillation, and it leads the open-weight tier on the first two. Whether that portfolio beats a rival's single stronger coding score is a per-team calculation, and the license conditions tilt it further for teams that want to train derivatives.

Why the release matters beyond the benchmarks

Three things make this release a marker rather than just another checkpoint.

It puts a US lab back in the large open-weight tier that Chinese labs have dominated through 2026, with Muse Spark Open bringing native audio and video input at a scale few open models offer. Whatever the benchmark gaps, the availability of a frontier-class, multimodal, million-token open-weight model from a US company changes the procurement conversation for organisations that want open weights but not a single vendor geography.

It demonstrates the unified-memory deployment path with Apple, two Mac Studios as one inference machine, which is a preview of where local model hosting goes as unified memory capacities climb. The two-vendor collaboration on an 816B model is the kind of thing that reshapes assumptions about what consumer-adjacent hardware can host.

And the safety-first framing of an open-weight release, with additional post-training specifically for standalone deployment and published refusal rates alongside capability numbers, sets a pattern worth watching. The model card states that a full safety evaluation report is coming, and that promise is worth holding Meta to, because the StrongREJECT number is the kind of claim that deserves independent replication.

For teams following the open-weights thread on this site, from the DeepSeek release to the harness economics, Muse Spark Open is the newest and largest entry, with the clearest self-reported trade: leading the tier on agency and safety, trailing it on code. The next data point that matters is not another Meta number but an independent Terminal-Bench run, because that is where the decision between this checkpoint and its rivals will actually be made.

Where the numbers stop

  • Every benchmark figure comes from Meta's model card, run at xhigh effort. No independent replication existed at publication, and Meta's own card says a full evaluation report is still coming.
  • The comparison set is Meta's choice of competitors at specific effort settings. Different effort levels or harnesses can reorder close scores, and the coding gaps, while consistent, have not been reproduced outside Meta.
  • The two-Mac-Studios deployment depends on hardware whose 512GB configuration had no announced price or ship date at release; cost figures in circulation, such as the four-machine estimate, are press estimates, not Meta list prices.
  • The license summary here is a reading, not legal advice. The Muse Model License and usage policy are the governing documents, and the derivative-works clause in particular should be reviewed by counsel before building products on distilled outputs.
  • The safety refusal rates are Meta's own adversarial evaluations, including internal benchmarks that cannot be independently checked. The StrongREJECT v2 figure is on a public benchmark and is the more verifiable of the claims.

FAQ

What is Muse Spark Open?

It is Meta's open-weight, natively multimodal mixture-of-experts model, released on 15 October 2026: 816 billion total parameters with about 42 billion active per token, a 1M-token context window, and input support for text, images, audio and video. It is built for coding and agentic work, ships in BF16 and NVFP4 checkpoints, and is optimized to run locally on two Mac Studios with M5 Ultra chips under the Muse Model License, with API access available through seven providers.

How does Muse Spark Open compare to GLM 5.3 and Kimi K3?

By Meta's own published tables, Muse Spark Open leads on agentic tool use and safety: 89.2 on MCP-Atlas and the top JobBench score against both, and a 0.5 score on StrongREJECT v2 adaptive attacks where the competitors score in the 60s and 80s. It trails on coding and general reasoning: 76.4 on Terminal-Bench 2.1 against 85.0 for Kimi K3, 60 on DeepSWE 1.1 against 68, and an overall index score of 39 against 45 and 44. The comparison is Meta's own at xhigh effort and awaits independent replication.

Can you actually run it locally?

Yes, but not cheaply. The published paths are two Mac Studios with M5 Ultra chips, each contributing up to 512GB of unified memory, or GPU deployments requiring at least 700GB of VRAM for the NVFP4 checkpoint and 2TB for BF16. Quantized versions run in around 344GB on a single 512GB machine. For most teams, hosted API access through the listed providers is the realistic option, and the local path matters as a demonstration of where unified memory hardware is heading rather than as a mainstream setup.

The short version

Muse Spark Open is Meta's 816B-parameter, 42B-active, 1M-context open-weight multimodal model, and the most credible US entry yet in the large open-weight tier. Meta's own numbers show it leading that tier on agentic tool use and safety refusals while trailing Kimi K3 and GLM 5.3 on coding, its stated training focus, a trade stated honestly in the model card. The two-Mac-Studios local path is real and historically significant for unified memory, but it costs as much as a car, so most adopters will take the API or a GPU node. The license is permissive with conditions: mandatory attribution, a 20-million-dollar MaaS agreement trigger, and broad derivative definitions that include distilled models. The release reshapes the open-weights conversation even where it does not win the benchmarks, and its claims now wait on the independent evaluations that its own card calls for.

Sources
Subscription Form