Meta has released Muse Spark Open, an open-weight model built for coding and agentic work that the company says can run locally on two Apple M5 Ultra Mac Studios. The headline numbers are substantial: 816 billion total parameters, 42 billion active at a time and a stated one-million-token context window. But the release is more precise than the marketing shorthand. The weights are available, yet the license is restrictive; local use is possible, yet it demands unusual memory and hardware; and the strongest performance claims come from Meta's own evaluations. The result is a consequential release that deserves both attention and qualification.
What Meta released
Meta announced Muse Spark Open on October 15 as a natively multimodal model aimed at long-horizon tasks, complex software engineering and multi-step reasoning. Its stated context window of one million tokens is designed to let a system retain far more material in a single interaction than a conventional chat prompt. Reuters independently reported the same core specifications and described the release as expanding developers' ability to download and adapt the technology, rather than keeping it behind a proprietary service.
The distinction between total and active parameters matters. Muse Spark Open is described as a mixture-of-experts Transformer with a perception encoder. The model card lists 816 billion parameters in total but 42 billion active per token, meaning routing selects a subset of experts for each step instead of calculating through the entire network every time. That design can reduce per-token computation relative to a dense model of the same total size, but it does not make the model small or easy to operate.
Multimodal by design
The model card lists text, images, video and audio as inputs, with text as the output. It says the training mix drew on text, images, audio and video from public sources, internet crawling, third parties, Meta products and services, and synthetic or externally curated material. Those descriptions establish the model's intended breadth, not proof that it performs equally well in every language or media format. The card itself warns that performance can decline outside strongly supported languages and that short, low-resolution video can be difficult to interpret.
Where developers can get it
Meta says weights are downloadable through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. It also lists hosted inference through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The announcement names support for SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth, with customization through PyTorch. In practical terms, that creates two paths: pay a provider for inference, or assume the engineering, memory and operational burden of running the checkpoint yourself.
Open-weight, not unrestricted open source
Calling Muse Spark Open open-weight is accurate; calling it unconditionally open source would erase a material qualification. Open weights let developers inspect, download and adapt trained parameters, unlike a closed model that remains controlled through an API. Reuters noted the potential advantages of that arrangement, including more control over data and a lower-cost alternative to some proprietary systems. Those benefits depend on the deployer's infrastructure and on what the license permits.
Meta says the release uses the Muse Model License with limited commercial terms. The model card says use is subject to a usage policy and other restrictions, including a separate-agreement requirement for certain Model-as-a-Service providers whose aggregate gross revenue exceeds $20 million during a consecutive 12-month period. That is not a minor footnote for companies planning to build a commercial hosted product. Legal and compliance teams would need to read the license and policy for their exact deployment, rather than infer permissions from the word open.
The same caution applies to claims of transparency. Access to weights can help researchers and developers study behavior, tune the model and keep sensitive workloads closer to their own environment. It does not expose every part of the training process, guarantee reproducibility, or remove the risk of inaccurate, biased or unsafe output. The model card recommends system-level guardrails, particularly when the model can call tools or make changes on a user's behalf. Downstream modifications can also change the safety profile Meta measured for the released checkpoint.
The two-Mac proposition
Meta's most vivid demonstration is local inference on two Apple M5 Ultra Mac Studios using Apple's MLX framework. The announcement says each machine's integrated GPU can use up to 512 GB of unified memory and cites 1.2 terabytes per second of memory bandwidth. In Meta's description, the pair behaves like one system with roughly a terabyte of unified memory. TechCrunch likewise reported that local operation requires at least about a terabyte of RAM and that the model is intended for the highest-memory Mac Studio configuration.
That is a technical milestone, not a claim that Muse Spark Open runs on an ordinary desktop. The official model card lists at least 700 GB of GPU VRAM for its NVFP4 version and 2 TB for BF16, in addition to a deployment framework. Quantization can reduce memory needs, but it may produce small quality differences in edge cases. A two-computer setup also brings cost, power, networking and maintenance considerations that are absent from a hosted API call.
For organizations with suitable Apple hardware, local execution could offer a different privacy and latency trade-off: data need not be sent to a third-party inference endpoint, and a team can control the serving stack. It can also make failures the team's responsibility. Memory capacity is only one constraint; sustained throughput, model loading, inter-machine coordination, updates and guardrails all affect whether a local installation is useful for production work.
Coding and agentic work
Meta positions the model less as a general chatbot than as an engine for software and tool-using workflows. The company says it paired Muse Spark Open with Muse Code and used it to make iOS changes to Ice Cubes, an open-source Mastodon client, including bug fixes, new features and a full build. Meta also says it drove DaVinci Resolve end to end, reviewing footage frame by frame and selecting crops for vertical video. These are demonstrations reported by Meta, not independent audits of reliability across real development teams.
The model card reports strong results on several agentic and coding benchmarks, including JobBench, MCP-Atlas, OSWorld Verified, DeepSWE 1.1 and Terminal-Bench 2.1. It also shows that Muse Spark Open does not lead every listed comparison: other models score higher on some coding, reasoning and multimodal tests. Benchmark results depend on the harness, tools, prompting, reasoning effort and model versions. The card says Muse Spark Open was evaluated with an xhigh reasoning setting, so a headline score should not be read as a universal measure of day-to-day engineering output.
Agentic capability changes the risk surface as well as the productivity promise. A model that can edit files, run commands or operate software can create useful automation, but it can also make an expensive mistake at machine speed. Meta reports safety testing and says the checkpoint met its release thresholds, while the model card stresses that evaluations do not cover every downstream configuration. Developers still need permissions, review points, isolation and logging around tools. Open weights make those controls more configurable; they do not make them optional.
What the evidence shows
The independently reported facts are clear: Meta released downloadable weights on October 15; the model has 816 billion total and 42 billion active parameters; it accepts multiple media types; and it can be accessed through third-party inference providers. Reuters and TechCrunch corroborate the Mac-local deployment story and the scale of the model. The less settled question is comparative superiority. Meta calls Muse Spark Open the most capable US open-weight model to date, but that is a company characterization supported by its selected evaluation methodology, not a consensus finding.
That distinction does not make the release unimportant. It shows how the frontier is moving in two directions at once: models are becoming large enough to require data-center-class resources, while hardware vendors and software stacks are trying to make a version of local operation practical. Muse Spark Open sits at that intersection. Its significance will be measured less by the launch slogan than by whether developers can use the weights within the license, afford the infrastructure and reproduce dependable results outside Meta's demonstrations.
FAQs
What is Muse Spark Open?
Muse Spark Open is Meta's open-weight, natively multimodal model for coding, agentic workflows and long-context reasoning. The released specification is 816 billion total parameters, 42 billion active parameters and a one-million-token inference context. Its inputs include text, images, video and audio, while its output is text, according to the model card.
Is it open source?
It is open-weight, meaning the trained weights can be downloaded and adapted, but it is not unrestricted open source. Meta distributes it under the Muse Model License, which includes a usage policy and limited commercial terms. A prospective deployer should review the license for its business model instead of treating the release label as a blanket permission.
Can it run on a Mac?
Yes, but the supported local configuration is demanding: Meta describes two Apple M5 Ultra Mac Studios running MLX. The model card separately lists at least 700 GB of VRAM for NVFP4 or 2 TB for BF16. That is specialized hardware, not a typical consumer Mac or ordinary desktop setup.
Are the benchmark results independently proven?
No. The model card documents Meta's comparisons and test settings, while Reuters and TechCrunch independently reported the release and its specifications. Those reports do not amount to a fresh third-party replication of every score. Results should therefore be treated as evidence under stated conditions, not a guarantee of superiority in every coding or agentic workload.