
Quick Answer
Meta released Muse Spark Open on October 15, 2026, giving developers an open-weight, multimodal model with 816 billion total parameters, 42 billion active parameters and a 1M-token context window. The model is aimed at coding and agentic work and, according to Meta, can run on two Apple M5 Ultra Mac Studios, while hosted access is available through seven inference platforms.
What Does Muse Spark Open Bring to Client Website Teams?
Muse Spark Open was built for long-horizon tasks, complex software engineering and multi-step reasoning. According to Meta, it accepts text, image, video and audio input and holds a million tokens of context.
Meta also states that the model can plan a change, call tools, check results and correct a failed step, which are core parts of an agentic coding workflow.
For agencies, the practical questions are local hardware, hosted operating cost, license conditions and whether the model’s measured capabilities fit the project. Agencies should review the model’s size and license terms before it enters a client production workflow.
How Does the Model Run on Local Mac Studio Hardware?
Meta optimized Muse Spark Open for the latest MLX framework on Mac Studio with M5 Ultra. According to Meta, each Mac Studio can use up to 512 GB of unified memory, so two machines can hold an entire model, and the two Mac Studios behave like one supercomputer with a terabyte of unified memory.
Meta states that the M5 Ultra provides 1.2 TB/s of memory bandwidth and that the chip’s bandwidth is fast enough to read the small slice each token needs. For local deployments, Meta provides an MLX Adapter, while the model can pull its weights directly from Hugging Face.
Meta reported two on-device workflows using Muse Code powered by Muse Spark Open. One shipped iOS bug fixes, new features and a full build for the open-source Mastodon client Ice Cubes. The other drove DaVinci Resolve end to end, with the model reviewing footage frame by frame to find the strongest moments and crop them vertically.
Where Can Developers Access the Weights and API?
Developers can download Muse Spark Open weights from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta also lists SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth as supported inference engines, with optimizations for AMD, Apple, Intel and NVIDIA silicon.
Hosted API access is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Meta’s developer documentation also lists support for BF16 and NVFP4 checkpoints, additional formats from hardware partners, and post-training options including TorchTitan, SGLang, NVIDIA NeMo AutoModel and Hugging Face TRL.
For a client project using hosted inference, LLM Stats reports prices of $1.25 per million input tokens and $4.25 per million output tokens via Together, with reused prompt prefixes costing $0.150 per million cached input tokens, while The AI Rankings lists $1.25 per million input tokens and $4.25 per million output tokens as the median host price, according to Artificial Analysis. Agencies should confirm the applicable terms before estimating a website project’s running cost.
What Do Independent Benchmarks Show?
According to LLMBenchmark.ai, Meta’s launch materials report 21 benchmark results, including 89.0% on GPQA Diamond, 58.5% on SciCode, 80.4% on OSWorld-Verified and 76.4% on Terminal-Bench 2.1 with the Terminus 2 agent. LLMBenchmark.ai notes that these are Meta’s reported figures rather than results it reproduced.
Independent results tell a more mixed picture. LLMBenchmark.ai recorded 95.0% on SWE-Bench Pro V2, ranking the model 12th among 17 models on that board, and 58.4% on SciCode, ranking it 17th among 111 models. According to LLMBenchmark.ai, its coding board also placed Muse Spark Open 41st of 106 on Vibe Code Bench with a score of 76.2% and 50th of 76 on code migration with a score of 21.2%.
On reasoning, LLMBenchmark.ai lists 28.0% on Humanity’s Last Exam at xhigh, 12th among 19 models in that setup, and 43.9% under an Artificial Analysis run, 30th among 189 models. According to ARC Prize, the team recorded 89.0% on ARC-AGI-1 and 57.5% on ARC-AGI-2 at xhigh reasoning, but 0.1% on ARC-AGI-3 with the Standard harness and 0.4% with the Provider Adapter harness. ARC Prize’s results show that the two ARC-AGI-3 harnesses produced sharply different scores.
Before using the model for client code, an agency can select the benchmark that matches its work and check the harness, reasoning setting and model version used for the score.
What Performance Does SGLang Report on Blackwell GPUs?
The SGLang team built a custom megakernel for Muse Spark Open and reported decoding speed increasing from 517 to 1,109 tokens per second on Blackwell GPUs. The SGLang team attributes the improvement to combining each layer’s attention and MoE work into one kernel and overlapping memory operations.
The post also describes the model’s architecture: 62 decoder layers, with 46 using sliding-window attention across the previous 2,048 tokens and 16 using global attention. Each token is routed to eight of 256 experts, and the team splits the model across eight GPUs with tensor parallelism.
Meta also reports that its custom, auto-tunable Helion kernels in SGLang deliver kernel-level speedups ranging from 1.03 to 3.95 times on NVIDIA and AMD GPUs, plus a 3% end-to-end improvement on AMD MI350. Agencies should compare these measurements against the hardware and serving stack planned for the client project.
What Does the Safety Report Say?
The 159-page Muse Spark Safety Preparedness Report, authored by Cristina Menghini and 118 other researchers, says Meta evaluated the model across Chemical and Biological, Cybersecurity and Loss of Control risks under its Advanced AI Scaling Framework. The report found elevated risks before mitigations, including Chemical and Biological capabilities assessed as likely reaching the high-risk category.
According to the report, Meta applied a multilayered set of mitigations and found that Muse Spark demonstrated state-of-the-art refusal across benchmarks for hazardous chemistry and biology workflows. The research assessed residual risk from deployment within Meta AI as acceptable under the framework.
The full evaluation report was still described as forthcoming in Meta’s release announcement.
What License Conditions Apply to Client Work?
Muse Spark Open is distributed under the Muse Model License with limited commercial terms. Meta says these terms allow many developers to use the model for free. LLM Stats says the license permits use, copying, modification, fine-tuning, derivative works and distribution, subject to attribution and the Usage Policy.
The LLM Stats license summary says licensees with aggregate gross revenue, including affiliates, above $20 million in any consecutive 12-month period must notify Meta and enter a separate written agreement before offering the software or derivatives as a Model-as-a-Service. It also says internal use, research and non-MaaS commercial product features are exempt from that threshold. Agencies should verify the applicable license terms.
An agency should review the license against the intended client use, distribution model and service structure.
What to Do Now
- List the client workflows that need coding, long documents, images, video or audio.
- Match each workflow to its benchmark, reasoning setting and harness.
- Compare hosted token rates with the cost of the intended Mac Studio or GPU deployment.
- Review the Muse Model License against the planned client use and distribution model.
- Record the target inference engine, hardware configuration and model checkpoint.
Key Considerations
- Meta says two M5 Ultra Mac Studios can hold an entire model, with each Mac Studio using up to 512 GB of unified memory.
- Meta describes Muse Spark Open as open-weight, while the Muse Model License includes limited commercial terms and a revenue threshold for Model-as-a-Service use.
- Independent benchmark positions vary by task, reasoning level and harness, so a general ranking does not establish fitness for a client workflow.
- Meta said the full evaluation report would become available soon, while detailed model-card results were already available at release.
FAQ
What Is Muse Spark Open?
Meta describes Muse Spark Open as its natively multimodal open-weight model with 816 billion total parameters, 42 billion active parameters and a 1-million-token context window. Meta describes it as built for coding and agentic work.
Can Muse Spark Open Run on a Desktop?
Yes. Meta optimized it to run on two Apple M5 Ultra Mac Studios, each with up to 512 GB of unified memory, using the latest MLX framework.
Where Can Developers Download Muse Spark Open Weights?
The weights are available from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.
Where Is Muse Spark Open Available Through an API?
Hosted API access is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Can Muse Spark Open Be Customized?
Yes. Meta says developers can customize the model with PyTorch and TorchTitan. Its developer documentation also lists additional post-training tools and frameworks available at release.
Related Coverage
- Thinking Machines Inkling 2 Small Claims Top Spot Among Western Open-Weight Models (November 6, 2026): Thinking Machines has released Inkling 2 Small, giving organizations a mixture-of-experts model with 594 billion total parameters and 17 billion active parameters.
- Meta’s Muse Wonder: A Frontier Model for Personal AI Agents (October 13, 2026): Businesses weighing whether to let an AI assistant handle real work now have a new model to assess: Meta’s Muse Wonder and the larger Muse Wonder Full, released on October 12, 2026.
Sources
Do you want help with your SEO or website development? Then Contact Us Here.
NewSunSEO
Grow in Google and AI Search
AI-SEO, local search, paid ads and websites from one team, with offices on Long Island and in Myrtle Beach.
- AI-SEO
- Local AI-SEO
- AI Consulting
- PPC
- Web Design
- Web Development
- Long Island Office(631) 320-1700
- Myrtle Beach Office(843) 492-7131
Request a Free Proposal
Tell us about your business and where you want to grow.
We will reply by email to talk through your goals and what a proposal would include.