🔍 Read the full analysis: Welcome Muse Spark Open: Open-weight Frontier Multimodal Mixture-of-experts Model From Meta on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Meta has released Muse Spark Open, an open-weight multimodal model aimed at agentic coding. Its model description lists 816 billion total parameters, a context window of up to one million tokens and substantial hardware requirements; independent performance results and the applicable license were not provided in the supplied material.
Meta has released Muse Spark Open, an open-weight model designed for multimodal, agentic coding, according to the model description on Hugging Face. The model accepts text, images, audio and video, and is listed at about 816 billion total parameters with a context window of up to one million tokens. Its scale means running it still requires substantial memory and multi-GPU infrastructure.
The model uses a mixture-of-experts architecture: about 42 billion parameters are active for each token, while the full 816 billion parameters remain part of the model. The description says each mixture-of-experts layer contains 256 experts, with a router selecting eight for each token. Muse Spark Open takes multimodal input but produces text, allowing developers to provide materials such as screenshots, voice notes or screen recordings alongside a coding task.
Meta describes the model as trained for developer workflows including reading codebases, refactoring across multiple files and extended debugging. The company also describes an agentic approach in which the model plans actions, observes results and revises its work over longer tasks, including web browsing with images. These are descriptions of intended capabilities; the supplied source does not include independent evaluations showing how well the model performs those tasks.
The published specifications list BF16 and NVFP4 checkpoints. BF16 files are about 1.5 terabytes in safetensors format. The source says running the BF16 version with Transformers uses four nodes, each equipped with eight H200 GPUs, while the NVFP4 version is designed for one B200 node. The source describes NVFP4 as having negligible accuracy loss, but provides no measurements or test conditions to support that claim.
The Hardware Cost of Open Weights
The release gives developers access to a model intended to combine coding-agent workflows with image, audio and video input. In principle, that could let a coding assistant consider a visual bug report or a recorded walkthrough together with a codebase, rather than relying only on text descriptions. Whether that approach improves results in real development work remains untested in the evidence provided.
The difference between active parameters and total parameters is central to deployment. Activating about 42 billion parameters per token may reduce the computation needed for each token relative to activating the whole model, but it does not mean the full model fits in memory as a 42-billion-parameter download. The source also estimates that a one-million-token context requires about 266 gigabytes for the BF16 key-value cache, apart from memory for the weights. Those requirements put practical evaluation beyond the reach of many ordinary local setups.
Open weights can give teams the option to inspect, run and potentially fine-tune a model, subject to its license and terms. But the supplied information does not state those terms or provide comparative coding results. Developers therefore have specifications and deployment recipes to investigate, not enough evidence to conclude that Muse Spark Open is more accurate, reliable or cost-effective than alternatives.
high performance GPU for AI training
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A Larger Model in Meta’s Muse Series
Muse Spark Open follows Muse Glimmer, which the source describes as a smaller, dense model intended for local use and released in August. Spark is the larger model in the series and uses a mixture-of-experts design with multimodal input. The source says the two models share a tokenizer, vocabulary and chat-template lineage; it does not give a release year for Glimmer or a date for Spark Open.
The model description lists a 62-layer language model with a hidden dimension of 8,192, alongside separate image and video perception and audio encoders. Images and video frames are processed by the perception encoder, while audio is converted into tokens at a stated rate of 12.5 per second. The one-million-token context limit applies to text and multimodal tokens together, according to the description. These architecture details describe how inputs are represented, not how accurately the model understands them.
The source points developers to checkpoint repositories named Muse-Spark-Open-816B-A42B and Muse-Spark-Open-816B-A42B-NVFP4. It also lists running and fine-tuning materials for Transformers, llama.cpp, vLLM and SGLang. Availability of those resources does not by itself establish that every framework supports every checkpoint or deployment configuration equally.
“The model has about 816 billion total parameters, with roughly 42 billion active per token.”
— Muse Spark Open model description on Hugging Face
large memory GPU for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Performance and License Remain Open
The supplied material does not give independent benchmark results, comparative coding scores or evidence about reliability on long-running tasks. Meta’s descriptions of intended training and capabilities are not independent proof that the model completes those tasks well or outperforms other coding models. There are also no quantified results for quality, latency or cost in multimodal coding workflows.
The source does not state the model’s license or full use and redistribution terms, nor does it provide a release date. It also gives no test data behind the stated NVFP4 accuracy trade-off. Those gaps matter for organizations deciding whether they may deploy or modify the weights, and for teams comparing precision options or estimating operating costs.
Hardware figures are presented as deployment guidance, not an independent audit of minimum requirements. Actual needs can vary with software, context length and workload. Although the source gives an estimate for the BF16 cache at a one-million-token context, it does not establish that every configuration can use that full context efficiently.
multimodal AI development hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Testing on Real Coding Workloads
Developers can use the listed checkpoint repositories and deployment recipes to assess installation requirements, run the model and explore fine-tuning. A useful next step for teams will be testing representative tasks, including multi-file changes and visual bug reports, while measuring accuracy, reliability, latency and hardware use against other models.
Further assessment will depend on information not included in the supplied description: the applicable license, reproducible evaluations and clearer test conditions for the NVFP4 claim. The source does not announce a separate next milestone or schedule. Until additional evidence or terms are available, the release establishes access to model weights and technical specifications, but not a verified performance ranking or a settled picture of its practical cost.
multi-GPU server for deep learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Muse Spark Open?
Muse Spark Open is an open-weight multimodal model from Meta, described as designed for agentic coding. It accepts text, images, audio and video and generates text.
How large is the model?
The model description lists about 816 billion total parameters, with roughly 42 billion active per token. The smaller active count does not remove the need to store the full model weights for deployment.
Can it run on a typical personal computer?
The deployment figures indicate substantial hardware needs. The source describes BF16 inference with Transformers using four nodes of eight H200 GPUs each, and the NVFP4 checkpoint as designed for a single B200 node; it does not describe a typical consumer-PC setup.
Has Muse Spark Open been shown to outperform other coding models?
No comparative results are included in the supplied material. Claims about coding and agentic capabilities are descriptions of intended use, not independent evidence of superiority.
What is still unknown about access to the weights?
The supplied source does not state the license or complete use and redistribution terms. Developers should check the repository terms before deciding how they can use or modify the checkpoints.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
