Meta released Muse Spark Open on October 15, 2026, an open-weight AI model the company is positioning as the most capable US open-weight system to date. The twist that’s driving the search traffic isn’t the benchmark score, it’s the hardware story: Meta says the model is built to run locally on two Apple M5 Ultra Mac Studio machines using the latest MLX framework, a claim first reported by 9to5Mac and echoed across The Verge, MacRumors, and TechCrunch.
That’s a notable pivot. Frontier-class open models have mostly been framed around data-center GPU clusters, not consumer-adjacent workstations sitting under a desk. By shipping checkpoints that fit inside a pair of Mac Studios, Meta is making an explicit bet that local, on-device AI compute matters enough to engineer for it directly, rather than leave it to third-party quantization hobbyists.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Muse Spark Open actually is
Muse Spark Open is an open-weight, natively multimodal mixture-of-experts model, according to Meta’s announcement and a model card published on Hugging Face. It accepts text, images, audio, and video as input and produces text output. Under the hood, the model routes tokens through 256 experts per layer, activating 8 of them per token, a routing pattern that keeps inference cost manageable while still drawing on a huge total parameter pool.
The headline numbers are big. Muse Spark Open carries roughly 816 billion total parameters, with about 42 billion active per token. That places it firmly in frontier mixture-of-experts territory, architecturally closer to the largest open releases from Chinese labs than to the leaner, single-digit-billion-parameter models built for phones. Meta ships two official checkpoints: Muse-Spark-Open-816B-A42B in BF16 precision, and Muse-Spark-Open-816B-A42B-NVFP4, a lower-precision variant aimed at cutting memory footprint without gutting quality.
Context length is set at up to 1 million tokens, enough to hold very long codebases, transcripts, or multi-document research bundles in a single pass. For developers who have been chaining retrieval pipelines to work around shorter context windows on rival open models, that’s a meaningful quality-of-life jump, assuming the effective recall at that length holds up under independent testing.
The Mac Studio angle: two M5 Ultra machines, 512GB each
Here’s the part driving the hardware story. Meta’s own claim, cited across the coverage, is that Muse Spark Open is “the first model in its class” able to “run on your desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework.” Reports describe a memory specification of up to 512GB per Mac Studio, which means the pair of machines combined can hold a model built around roughly 816 billion parameters, something that would otherwise require a rack of data-center GPUs.
MLX is Apple’s own machine-learning framework, tuned for the unified-memory architecture that Apple Silicon uses instead of a traditional discrete-GPU-plus-system-RAM split. Because a Mac Studio’s memory is addressable by both CPU and GPU cores without a slow PCIe hop, splitting a huge mixture-of-experts model across two machines becomes a software problem Apple’s own tooling is built to handle, rather than a from-scratch engineering project for Meta’s research team.
What’s still unconfirmed is pricing. Neither Meta’s announcement nor the outlets covering it as of October 15 nailed down what a two-Mac-Studio Muse Spark Open setup costs end to end, and nobody has published an official number for the Mac Studio or M5 Ultra configuration required. Anyone shopping for the hardware this week should treat specific dollar figures circulating online as speculation until Apple or Meta confirms a configuration.
Why running a frontier model on a desktop matters
Most frontier-scale open-weight releases over the past two years have assumed the end user has access to multi-GPU clusters, whether that’s a rented cloud instance or an on-prem NVIDIA deployment. Meta’s framing flips that assumption for Muse Spark Open. If a studio, research lab, or well-funded engineering team can run a ~816-billion-parameter multimodal model on two prosumer-adjacent workstations instead of a server room, the cost and complexity of experimenting with frontier open weights drops by an order of magnitude.
That matters especially for workloads where sending data to a third-party API isn’t an option, legal teams, security-sensitive engineering groups, or any organization running under strict data-residency rules. A model that fits on hardware physically sitting in a locked room removes an entire category of vendor-risk conversation that comes with calling a hosted inference endpoint.
It also puts pressure on NVIDIA’s narrative that serious open-weight inference requires its hardware stack. Meta hasn’t abandoned NVIDIA hardware, in fact its own technical blog post describes scaling the same model’s agentic coding workflows on NVIDIA’s GB300 NVL72 platform for production-scale throughput. But by simultaneously showing the model runs on two Mac Studios, Meta is signaling that the same weights scale down to a desk as well as up to a data center, which is a flexibility story none of its direct open-weight competitors have made as explicitly.
Licensing: what “limited commercial use” means
Muse Spark Open ships under Meta’s own Muse Model License, which the company describes as permitting limited commercial use. That’s consistent with how Meta has licensed its other open releases: broad enough to let startups and researchers build on the weights, but with guardrails that keep Meta’s legal team comfortable with downstream deployment at scale. The exact commercial thresholds and restrictions in the Muse Model License weren’t fully detailed in the initial coverage, so teams planning to ship a commercial product on top of Muse Spark Open should read the license text on Hugging Face directly before committing engineering time.
Licensing terms like this have become one of the more contested fronts in the open-weight race. A model that’s technically downloadable isn’t necessarily “open” in the sense a startup’s legal counsel would sign off on for a commercial product, and Meta’s history with the Llama license family shows the company tightening and loosening commercial terms release over release.
Muse Spark Open at a glance
| Spec | Muse Spark Open |
|---|---|
| Total parameters | ~816 billion |
| Active parameters per token | ~42 billion |
| Architecture | Mixture-of-experts, natively multimodal |
| Experts per layer / selected per token | 256 / 8 |
| Context window | Up to 1 million tokens |
| Inputs | Text, images, audio, video |
| Output | Text |
| Checkpoint formats | BF16, NVFP4 |
| Official checkpoint names | Muse-Spark-Open-816B-A42B, Muse-Spark-Open-816B-A42B-NVFP4 |
| Local hardware target | Two Apple M5 Ultra Mac Studio systems, up to 512GB memory each, via MLX |
| Distribution | Hugging Face |
| License | Muse Model License (limited commercial use) |
| Release date | October 15, 2026 |
How Muse Spark Open stacks up against other open-weight models
The open-weight field has gotten crowded, and Meta is entering it against labs that have each taken a different scaling bet. DeepSeek’s V3 architecture popularized the mixture-of-experts approach at scale with roughly 671 billion total parameters and about 37 billion active per token. Alibaba’s Qwen3 line includes a 235-billion-parameter mixture-of-experts variant with about 22 billion active. Meta’s own earlier Llama 4 family split into a 400-billion-parameter Maverick model with 17 billion active, and a smaller 109-billion-parameter Scout variant built for longer context on less hardware.
Against that backdrop, Muse Spark Open’s roughly 816 billion total parameters make it one of the largest open-weight releases to date by total parameter count, while its 42-billion active-parameter figure keeps inference cost in a similar range to its mixture-of-experts peers. The real differentiator isn’t the parameter count on a spec sheet, though, it’s that Meta is the only one of these labs publicly attaching a specific consumer-adjacent hardware path (two Mac Studios) to a model this size.
| Model | Developer | Total params | Active params | Primary local hardware path |
|---|---|---|---|---|
| Muse Spark Open | Meta | ~816B | ~42B | 2x Apple M5 Ultra Mac Studio (MLX) |
| DeepSeek V3 | DeepSeek | ~671B | ~37B | Multi-GPU server clusters |
| Llama 4 Maverick | Meta | ~400B | ~17B | Multi-GPU server clusters |
| Qwen3-235B-A22B | Alibaba | ~235B | ~22B | Single high-memory GPU or server node |
| Llama 4 Scout | Meta | ~109B | ~17B | Single high-memory GPU |
Meta’s own framing leans into this gap. The company’s claim that Muse Spark Open is “the most capable US open-weight model to date” is explicitly a response to the pace Chinese labs have set in open-weight releases over the past year, a point The New Stack made directly in its coverage, describing the release as Meta’s answer to China’s leading open-weight efforts.
The competitive subtext: catching up to Chinese open models
For most of 2025 and into 2026, the loudest open-weight headlines came out of Chinese labs. DeepSeek’s releases forced a broader rethink of training efficiency, and models from Alibaba’s Qwen team and Zhipu’s GLM family kept pushing open benchmark leaderboards. US labs answered mostly with closed, API-gated frontier models, Anthropic’s Claude line, OpenAI’s GPT series, and Google’s Gemini family all stayed proprietary even as their capabilities climbed.
Muse Spark Open is Meta doubling down on the opposite strategy at a moment when that strategy has mostly been ceded to Chinese developers. By pairing a large, genuinely open checkpoint with a hardware story built around Apple Silicon rather than NVIDIA racks, Meta is trying to make the release relevant to a US-based developer and enterprise audience that might not otherwise have touched a Chinese-origin open model for data-governance or procurement reasons.
It’s also consistent with Meta’s broader AI roadmap, which has increasingly split into a closed “Muse” line of in-house products and an open line meant to seed developer ecosystems the way the original Llama releases did in 2023 and 2024. Muse Spark Open appears designed to recapture some of that ecosystem-building momentum at the exact moment it had started shifting toward Chinese open-weight releases.
Historical context: from Llama to Muse
Meta’s open-weight strategy has evolved in clear stages. Llama and Llama 2 established the company as the default source for open base models developers could fine-tune without a research-lab budget. Llama 3 and Llama 4 scaled that approach up into mixture-of-experts territory, with Llama 4’s Maverick and Scout variants representing Meta’s first major MoE open releases at hundreds of billions of parameters.
Muse Spark Open represents a further jump, both in raw scale (816 billion total parameters versus Llama 4 Maverick’s 400 billion) and in hardware ambition. Where Llama 4 models were primarily designed around GPU-cluster deployment, Muse Spark Open is the first release where Meta has publicly engineered a consumer-workstation hardware path as a first-class deployment target, not an afterthought left to the open-source quantization community.
Why Apple’s MLX framework matters here
Apple introduced MLX as an array-computing framework optimized for its own chips, and it has steadily become the go-to toolkit for running large models locally on Apple Silicon. The framework’s design around unified memory is exactly what makes splitting an 816-billion-parameter model across two physical Mac Studios tractable, since large tensors can be sharded across machines without the kind of PCIe bandwidth bottleneck a traditional discrete-GPU setup would hit. Meta building official support for MLX, rather than leaving Apple Silicon compatibility to community projects, is itself notable: it suggests direct coordination between the two companies around this release, something CNBC reported explicitly, describing Apple and Meta as partnering to make the Muse model run on Mac computers.
What the NVFP4 checkpoint is for
The NVFP4 checkpoint variant exists for a different audience than the Mac Studio setup: it’s a lower-precision format designed to shrink memory and compute requirements on NVIDIA hardware that supports the FP4 numeric format, which is why Meta’s own developer blog walks through deploying Muse Spark Open on NVIDIA’s GB300 NVL72 platform. Running both a Mac-optimized BF16 path and an NVIDIA-optimized NVFP4 path from the same release lets Meta cover both the “run it on a desk” story and the “run it at data-center throughput” story with one model family.
Market impact: what this means for AI hardware demand
A frontier-scale open model that officially targets Apple Silicon, rather than treating it as an unsupported edge case, is a demand signal Apple has been trying to generate for years. The Mac Studio has always had workstation-class memory as a selling point, but most of the AI industry’s purchasing dollars have flowed toward NVIDIA GPUs, not Apple Silicon, because the software ecosystem and model releases were built around CUDA first. Meta building and publishing an official MLX path changes that calculus for at least one flagship model, and gives Apple a concrete example to point to when pitching Mac Studio as AI-capable hardware rather than just a high-end editing workstation.
For NVIDIA, the practical risk is limited in the near term, since the NVFP4 checkpoint and the GB300 NVL72 deployment path mean Meta isn’t steering enterprise-scale inference away from NVIDIA hardware. But symbolically, a two-Mac-Studio alternative to a GPU rack for running an 816-billion-parameter model chips away at the idea that NVIDIA hardware is the only serious way to run frontier-scale open weights, even if it remains the fastest way to run them at volume.
For the broader AI hardware market, the release adds to a year that’s already been defined by compute scarcity and pricing pressure, visible in everything from HBM4 memory shortages pushing prices up sharply to GPU price spikes rippling through consumer and enterprise channels alike. A credible desktop-scale path for frontier open models doesn’t fix that scarcity, but it does give a subset of developers and enterprises a way to sidestep it entirely for certain workloads.
What developers and enterprises should watch next
Several open questions will shape how much Muse Spark Open actually gets used outside of benchmark comparisons. First, independent benchmarking: Meta’s own claim of being the most capable US open-weight model needs to hold up against third-party evaluation, not just Meta’s internal numbers. Second, real-world Mac Studio performance: raw memory capacity to hold the weights is different from usable tokens-per-second throughput for interactive use, and early adopters running the two-Mac-Studio configuration will be the ones who surface whether it’s a genuinely usable local setup or a proof-of-concept that’s technically possible but impractically slow.
Third, the commercial license’s fine print will matter more as startups look at building products on Muse Spark Open rather than just experimenting with it. And fourth, how competitors respond: if Muse Spark Open’s Mac Studio story generates real developer interest, it would be a surprise if DeepSeek, Alibaba, or Zhipu didn’t follow with their own officially supported Apple Silicon deployment paths for their next major open releases, especially given how closely those labs have mirrored each other’s moves throughout 2025 and 2026.
Predictions for the months ahead
- Independent benchmark results on Muse Spark Open will surface within weeks of the October 15 release, and the real test will be whether its multimodal input handling (audio and video specifically) matches the quality of its text generation, not just raw parameter count.
- Expect Apple to lean on this release in its own marketing for Mac Studio and future Apple Silicon chips, since an officially-supported frontier-scale model is exactly the kind of proof point Apple has lacked in the AI hardware conversation.
- Rival open-weight labs, particularly DeepSeek, Alibaba’s Qwen team, and Zhipu’s GLM team, will likely publish their own MLX or Apple Silicon support within the next release cycle, following Meta’s lead rather than ceding that audience.
- Pricing details for a Mac Studio configuration capable of running Muse Spark Open will emerge from third-party retailers and reviewers well before Apple or Meta publishes anything official, given how unconfirmed the current cost picture is.
- NVIDIA will continue to dominate enterprise-scale deployment of Muse Spark Open through the NVFP4 and GB300 NVL72 path, even as the Mac Studio option draws developer attention; the two hardware stories are complementary rather than competing for Meta’s purposes.
How this fits the bigger AI model landscape
Muse Spark Open lands in a market where the pace of open and closed model releases has made it hard for any single system to hold the spotlight for long. Just this month, comparisons like Claude Opus 5.5 versus GPT-6 Astra and GLM 5.3 Prime versus Qwen3.8 Max versus Step 5 have shown how tight the gap between frontier labs has become, whether open or closed. Meta’s other 2026 release, Muse Wonder, took the closed, subscription-gated approach aimed directly at OpenAI’s consumer business. Muse Spark Open is the mirror image of that strategy: open weights, developer-first distribution, and a hardware story built around owning the model outright instead of renting access to it.
That dual-track approach, closed consumer product on one side, open developer-facing model on the other, mirrors how Chinese open-source LLMs have been reshaping AI infrastructure spending decisions throughout 2026. If Muse Spark Open succeeds at pulling developer mindshare back toward a US-origin open model, it could meaningfully shift where the next wave of fine-tuning, tooling, and enterprise deployment work gets built.
Frequently asked questions
What is Meta’s Muse Spark Open model?
Muse Spark Open is Meta’s open-weight, natively multimodal mixture-of-experts AI model released on October 15, 2026. It accepts text, images, audio, and video input and produces text output, with roughly 816 billion total parameters and about 42 billion active per token.
Can Muse Spark Open really run on a Mac Studio?
According to Meta’s announcement, Muse Spark Open can run locally using two Apple M5 Ultra Mac Studio systems with up to 512GB of memory each, using Apple’s MLX framework. Meta describes it as the first model in its class able to run on that configuration.
Do I need two Mac Studios, or does one work?
Reports describe the local setup as requiring two Mac Studios working together, with their combined memory needed to hold the full model. A single Mac Studio, even at 512GB, was not described as sufficient for the full checkpoint in the coverage of this release.
Where can I download Muse Spark Open?
Muse Spark Open is available through Hugging Face, where Meta published both the BF16 checkpoint and the NVFP4 variant along with a model card detailing the architecture and license.
Is Muse Spark Open free to use commercially?
The model ships under Meta’s Muse Model License, which permits limited commercial use. The exact terms and restrictions should be reviewed directly on the model’s Hugging Face license page before building a commercial product on top of it.
How does Muse Spark Open compare to DeepSeek V3 or Qwen3?
Muse Spark Open’s roughly 816 billion total parameters make it larger by total parameter count than DeepSeek V3 (about 671 billion) and Qwen3-235B-A22B (about 235 billion), though all three use a mixture-of-experts design that keeps active parameters per token far lower than the total. Muse Spark Open’s main differentiator is the officially supported Mac Studio deployment path, which none of the other models have matched.
Why did Meta build an NVFP4 checkpoint too?
The NVFP4 checkpoint is a lower-precision format aimed at NVIDIA hardware, letting Muse Spark Open run at data-center throughput on platforms like NVIDIA’s GB300 NVL72, according to Meta’s own developer blog. It covers enterprise-scale deployment while the BF16/MLX path covers local, desktop-scale use.
Does this mean Apple is getting into AI model development?
Not directly. Muse Spark Open is Meta’s model, not Apple’s. But CNBC’s reporting describes direct coordination between Apple and Meta to make the model run well on Mac hardware, which signals a hardware-software partnership even though Apple isn’t credited as a co-developer of the model itself.
