🔍 Read the full analysis: Inside Meta’s Muse Spark Open: A Frontier Mixture-of-Experts Model on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
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, about 42 billion active per token and a context window of up to one million tokens, but the supplied material does not include independent performance results or the model’s license.
Meta has released Muse Spark Open, an open-weight model designed for multimodal, agentic coding, according to the original analysis and its model description on Hugging Face. The model accepts text, images, audio and video, and is listed with 816 billion total parameters, roughly 42 billion active per token and a context window of up to one million tokens. Those specifications make it a substantial infrastructure commitment; the available material does not provide independent tests showing how well it performs.
Muse Spark Open is built to take on developer tasks such as reading codebases, refactoring multiple files and handling extended debugging sessions. Meta describes a workflow in which the model plans, takes actions, observes results and revises its work, including browsing the web with images. These are descriptions of intended capabilities, not independently verified results in the supplied material.
The model uses a mixture-of-experts architecture. Its layers contain 256 experts, with a routing system selecting eight for each token. While about 42 billion parameters are active for a given token, the full 816 billion parameters still need to be held in memory for deployment. The one-million-token context can include text and multimodal tokens, but using that capacity also adds substantial memory demands.
The release is listed in BF16 and NVFP4 checkpoint formats. The BF16 files total about 1.5 terabytes in safetensors format. The source says BF16 inference in Transformers uses four nodes with eight H200 GPUs each; the NVFP4 checkpoint is designed for a single B200 node. The source describes the four-bit version as having negligible accuracy loss, but supplies no measurements or test conditions to support that claim.
A Large Model for Multimodal Coding
The release brings open weights, multimodal inputs and a coding-agent focus together in a model aimed at tasks that can span many files and extended interactions. In principle, a developer could provide a screenshot, voice note or screen recording alongside a codebase, giving the model several forms of evidence about a software issue without first converting everything into text.
That flexibility may matter to teams investigating visual bugs or following recorded walkthroughs, but the practical value remains unmeasured in the supplied source. Accuracy, reliability and operating cost will determine whether the model helps with real development work. Its active-parameter count does not make it a lightweight local model: all 816 billion parameters remain part of the deployment memory requirement, before accounting for context and other system needs.
Open weights can give developers the ability to run and fine-tune a published model, subject to its terms. Here, however, the license and use conditions are not stated in the supplied material. Those details, along with reproducible comparisons against other coding models, will affect how organizations assess the release.
high performance GPU for AI training
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Muse Spark Fits the Series
Muse Spark Open is described as the larger model in Meta’s Muse series, following Muse Glimmer, which the source says was released in August as a smaller, dense model intended for local use. The source does not give a year for that August release. Spark instead uses a mixture-of-experts design and adds multimodal features while sharing a tokenizer and vocabulary with Glimmer.
The model description lists a 62-layer language model with a hidden dimension of 8,192, plus separate image and video perception and audio encoders. Images and video frames use the perception encoder; audio is converted into tokens at a stated rate of 12.5 per second. These are architectural specifications, not evidence of how accurately the model understands those inputs. The source estimates that the key-value cache for a full one-million-token context would require about 266 gigabytes in BF16, in addition to memory for the model weights.
The source points to checkpoint repositories under the names Muse-Spark-Open-816B-A42B and Muse-Spark-Open-816B-A42B-NVFP4. It also describes deployment recipes for Transformers, llama.cpp, vLLM and SGLang, as well as fine-tuning guidance. The supplied material does not state when the release occurred or provide full access terms.
“The NVFP4 version is described as having negligible accuracy loss.”
— The supplied source’s summary of the model description
Performance and Access Terms Remain Open
The supplied material includes no independent benchmark results, comparative coding scores or evidence about the model’s reliability on extended tasks. Meta’s descriptions of planning, tool use and multimodal coding should not be treated as proof that the model outperforms alternatives. The NVFP4 accuracy statement also lacks testing details in the material provided.
The model’s license, release date and full use and redistribution terms are not given. It is also unclear what performance and cost teams can expect outside the specified hardware setups, or whether the listed deployment recipes support every intended workload equally. The one-million-token limit is a stated maximum; the source’s memory estimate shows that using the full context may require substantial additional resources.
Developers Can Test the Checkpoints
Developers can use the listed checkpoints and deployment recipes to test Muse Spark Open on their own coding and multimodal tasks. Those tests can help teams measure quality, reliability, latency and hardware cost under their specific workloads. The supplied material does not announce a separate next milestone or provide a schedule for further evaluations.
Important details for broader assessment remain the applicable license, reproducible performance results and measured trade-offs between BF16 and NVFP4. Until those are available, the model’s specifications describe what has been released, while its practical performance and accessibility remain matters for evaluation.
Key Questions
What is Muse Spark Open?
It is an open-weight multimodal model from Meta designed for agentic coding tasks. 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 active count does not remove the need to hold the full model’s parameters in memory for deployment.
Can it run on a standard desktop computer?
The supplied hardware details indicate that this is a large infrastructure-scale model, not a typical desktop download. The source describes BF16 inference using four eight-H200-GPU nodes and an NVFP4 setup designed for one B200 node.
Has Muse Spark Open been shown to outperform other coding models?
Not in the supplied material. It provides no independent benchmarks or comparative coding scores; Meta’s capability descriptions are not independent performance evidence.
What are the model’s license terms?
The supplied material does not specify the license or full terms for use and redistribution. Developers should check the checkpoint repositories for the applicable terms before relying on the weights.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
