TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Meta says Muse Spark Open is available today as an open-weight model for coding, multimodal and agentic tasks. Its 816 billion total parameters, 1-million-token context window and support for hosted services and local hardware make it a substantial release, though Meta’s performance and safety claims rely on its own evaluations.
Meta has released the weights for Muse Spark Open, an open-weight model it describes as its most capable U.S. open-weight release to date, with 816 billion total parameters and a 1-million-token context window, following the earlier Muse Spark 1.3 release. The model is aimed at coding and agentic work, and Meta says it can run locally across two Apple M5 Ultra Mac Studios using the MLX framework, building on the coding focus discussed in Meta’s Muse Spark 1.2 launch.
Meta describes Muse Spark Open as natively multimodal, with 42 billion active parameters and a context window of 1 million tokens. The company says it is designed for long-horizon tasks, complex software engineering and multi-step reasoning. Those are Meta’s descriptions of the model’s intended strengths; the provided release does not include benchmark scores with which to independently compare them.
Developers can access the model through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Downloadable weights are listed on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models. Meta says the model works with inference engines including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for AMD, Apple, Intel and NVIDIA hardware.
For local use, Meta says two M5 Ultra Mac Studios can hold the model using their combined memory capacity, with an MLX adapter for running it on Apple Silicon. The company also reports testing on-device agent workflows through Muse Code, including changes to the open-source iOS app Ice Cubes and an end-to-end DaVinci Resolve workflow, similar to the comparative app-building work in tests of Muse Spark and other AI models. These examples are company-reported demonstrations, not independent evaluations of general performance.
Local Access and Developer Options
The release gives developers another option for using a large model outside a single hosted service. They can download the weights, run the model through multiple inference systems or use one of the listed API providers. That flexibility may help teams test deployment choices and adapt the model for particular workloads, while the actual cost and practicality will depend on hardware, serving requirements and licensing.
Meta’s two-computer desktop setup draws attention to the hardware demands behind local use. The company says each M5 Ultra Mac Studio can provide up to 512 GB of unified memory, allowing two systems to hold the model together. This is a specific configuration, not evidence that the model can run on an ordinary desktop. Meta also reports custom Helion GPU kernels in SGLang with 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs, and a 3% end-to-end speedup on AMD MI350. The figures describe different measurements and should not be treated as a single overall performance gain.
The licensing terms also affect how broadly the release can be used. Meta says the Muse Model License allows many developers to use the weights for free but includes limited commercial terms. Teams considering commercial deployment will need to review the license rather than assume the weights are unrestricted.
As an affiliate, we earn on qualifying purchases.
How Meta Is Distributing the Model
Meta is releasing model weights rather than limiting access to a single hosted product. The company presents that approach as a way for people to run and customize advanced models themselves. In this release, access spans cloud marketplaces, API providers and local inference tools, while PyTorch’s TorchTitan is named as an option for further tuning.
The release also includes a safety component. Meta says it applied additional safety post-training because open-weight releases can carry particular risks. It reports evaluating the model under its Superintelligence Scaling Framework and says the results met the framework’s release thresholds. Meta points readers to the model card for details and says a full evaluation report will be available later. These are the company’s stated findings, not an external certification.
““The most capable US open-weight model to date.””
— Meta
As an affiliate, we earn on qualifying purchases.
Benchmarks, Safety and Practical Costs
The announcement refers readers to a separate evaluation methodology report but does not supply benchmark scores in the provided material. It is therefore not possible from this release alone to verify how Muse Spark Open compares with other models on coding, reasoning or multimodal tasks. Meta’s examples of software and video workflows also do not establish how reliably the model performs across other projects or users.
Some release details remain open. Meta says a full safety evaluation report is forthcoming, and the supplied information does not specify its publication date. The announcement also does not give a calendar date for “available today,” or a complete breakdown of the costs, speed and hardware requirements for different deployment configurations. Those factors may vary across providers and systems.
As an affiliate, we earn on qualifying purchases.
Developer Access and Evaluation Reports
Developers can access the weights through the named model marketplaces and Hugging Face, or try hosted inference through Meta’s listed API partners. For local Apple Silicon use, Meta directs developers to the MLX adapter; for additional customization, it points to PyTorch TorchTitan. Availability and performance may differ by platform, hardware and provider.
The next source-backed milestone is Meta’s promised full evaluation report. Its publication should provide more detail on the methods and results behind the company’s benchmark and safety statements. Until then, independent testing and practical deployment reports will help clarify how the model performs beyond Meta’s own demonstrations.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Muse Spark Open?
Muse Spark Open is Meta’s open-weight multimodal model, described as having 816 billion total parameters, 42 billion active parameters and a 1-million-token context window.
Where can developers get it?
Meta lists downloadable weights on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted access is listed through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Can it run on a desktop?
Meta says it is optimized for a setup using two Apple M5 Ultra Mac Studios with MLX. That statement describes a high-memory configuration and does not mean the model will run on typical consumer desktops.
Is the model’s safety evaluation independently verified?
The announcement reports Meta’s own evaluation under its Superintelligence Scaling Framework and says the model met release thresholds. A full evaluation report is promised, but the source does not describe an independent audit.
Can companies use the weights commercially?
Meta says the Muse Model License permits many developers to use the model for free but has limited commercial terms. Prospective users should check the license for the conditions that apply to their use.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
