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 has released Muse Spark Open, an open-weight multimodal model designed for agentic coding. The model has about 816 billion total parameters, activates about 42 billion per token and supports a context window of up to one million tokens; deployment still requires memory for the full model.
Meta has released Muse Spark Open, an open-weight model designed for agentic coding that accepts text, images, audio and video. Its mixture-of-experts architecture has about 816 billion total parameters, with roughly 42 billion active for each token, and a context window of up to one million tokens, according to the model description published on Hugging Face.
The model is trained for developer workflows such as reading codebases, multi-file refactors and long debugging sessions. Meta describes it as able to plan, take actions, observe results and correct its work over longer tasks, including browsing the web with images. Those are descriptions of its intended capabilities; the supplied material does not provide independent evaluation results.
Muse Spark Open takes multimodal input but generates text. That gives developers a common interface for asking about different kinds of material: for example, a screenshot, voice note or screen recording related to a coding task. Its mixture-of-experts layers contain 256 experts apiece, with a router selecting eight for each token.
The model is available in BF16 and NVFP4 checkpoint formats. The BF16 files are about 1.5 terabytes in safetensors format. The source says BF16 inference in Transformers uses four nodes, each with eight H200 GPUs, while the NVFP4 version is designed for a single B200 node. The latter uses four-bit precision; the claim of negligible accuracy loss comes from the source description.
Model Brief · Meta · Open Weights
Welcome Muse Spark Open: Meta’s Frontier Multimodal Model
An infrastructure scale mixture of experts model built for agentic coding. It takes text, images, audio and video as input, then responds in text, with a context window of up to one million tokens.
Text and multimodal tokens
Eight selected for each token
Safetensors files
BF16 source estimate
One interface for varied coding inputs
Muse Spark Open combines multimodal perception with a model trained for long developer workflows. Its described capabilities are design intent, not independently verified results.
Images and video
Images and video frames pass through a perception encoder. A developer could ask about a screenshot, visual bug report or recorded walkthrough alongside a codebase.
Voice and sound
Separate audio encoding turns audio into tokens at 12.5 tokens per second, giving the model another way to receive task context.
One text response
The model accepts text, images, audio and video, but generates text. The shared interface may connect media with a coding task; practical gains remain unmeasured in the supplied material.
MoE efficiency, infrastructure scale
The active parameter count describes per token computation. It does not describe the memory needed to host the complete model.
How routing works
Each mixture of experts layer has 256 experts. A router selects eight for each token, while the full set of experts remains part of the deployed model.
Deployment footprint
Source deployment notes point to multi-node accelerators for BF16 and a single B200 node for NVFP4. Hardware needs can vary by inference software and workload.
The cache estimate is for a full one million token context in BF16, in addition to weight memory. Bar lengths are illustrative, not a shared scale.
A long task, in four stages
Meta describes the model as able to plan, act, observe and correct over longer tasks, including web browsing with images.
Gather context
Read code, a visual report, audio note or screen recording in the same session.
Map the task
Work through a codebase, multi-file refactor or extended debugging request.
Use tools
Take actions and inspect results as the task proceeds, according to the model description.
Correct course
Use observations to adjust the next step. Reliability on real tasks is not established here.
From Muse Glimmer to Spark
Spark is the larger, multimodal mixture of experts model in Meta’s Muse series. The release description also outlines its model structure and checkpoint options.
A larger step in the series
Muse Glimmer was described as a smaller, dense model intended for local use. Spark shares the tokenizer, vocabulary and chat template lineage, while adding multimodal features and an MoE design.
Language plus perception
The description lists a 62-layer language model with an 8,192 hidden dimension, plus separate image and video perception and audio encoders. These specifications do not establish interpretive accuracy.
Four accelerator nodes
The source says BF16 inference in Transformers uses four nodes, each with eight H200 GPUs. The published safetensors files are about 1.5 terabytes.
One B200 node
The four-bit NVFP4 checkpoint is specified for one B200 node. The source claims negligible accuracy loss but provides no measurements or test conditions here.
What the supplied material establishes
Specifications and release descriptions help frame an evaluation. They do not substitute for measured performance or clear usage terms.
| Question | What is stated | What remains open |
|---|---|---|
| Model scale | ✓ About 816B total and 42B active per token | Practical cost and throughput by workload |
| Context | ✓ Up to 1M text and multimodal tokens | Performance and memory under real use |
| Accuracy | ~ NVFP4 is described as having negligible loss | Measurements, test conditions and independent results |
| Coding ability | ~ Trained for long coding workflows | Benchmarks and comparative coding scores |
| Release terms | ✗ Not specified in supplied material | License, redistribution terms and release date |
A meaningful release to evaluate
Open weights, multimodal input and a coding agent focus make Spark relevant for teams with the infrastructure to test it.
Why it stands out
The combination of a long context, multiple input types and published checkpoints gives developers a basis for testing against real multimodal coding workflows.
Why adoption may be hard
A 1.5 TB BF16 checkpoint and a large context cache require substantial infrastructure. The supplied material includes no independent results to justify that deployment cost.
What would change the picture
Reproducible gains on long coding and multimodal debugging tasks, clear cost advantages, explicit licensing and measured NVFP4 accuracy would strengthen the case.
How teams can assess it
Use the listed deployment recipes to check hardware needs, then evaluate reliability, latency and cost on representative tasks. No separate next milestone is announced in the supplied material.
The short answers
Core facts for developers deciding whether to inspect the checkpoints and run their own evaluation.
What is Muse Spark Open?
A Meta open-weight model designed for agentic coding. It accepts text, images, audio and video, and generates text.
How large is it?
About 816 billion total parameters, with roughly 42 billion active per token. Deployment still needs memory for all weights.
What hardware is described?
BF16 Transformers inference: four nodes with eight H200 GPUs each. NVFP4: one B200 node. Actual requirements depend on software and workload.
Does this prove it leads coding models?
No. The supplied material contains no independent benchmarks or comparative coding scores. Capability descriptions are not proof of superior performance.
Multimodal Input for Coding Agents
Muse Spark Open brings several input types into a model aimed at long-running coding tasks. A developer could, in principle, ask it to relate a visual bug report or recorded walkthrough to a codebase without first converting every item into text. Whether this improves results in real workflows depends on performance, reliability and cost, none of which are quantified in the supplied material.
The distinction between active parameters and deployment memory matters for teams considering the model. Activating about 42 billion parameters per token may reduce computation compared with running every parameter for every token, but all 816 billion parameters remain in memory. That makes this a model for substantial multi-GPU infrastructure, rather than a lightweight local download, despite the smaller active count.
For developers, open weights offer the option to inspect, run and fine-tune the published checkpoints under their applicable terms. The source introduces instructions and recipes for running and fine-tuning the model, but the provided material does not specify its license or establish how it performs against other coding models.
high performance GPU for AI development
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Muse Glimmer to Spark
Muse Spark Open is the larger model in Meta’s Muse series. The source says Muse Glimmer was released in August as a smaller, dense model intended for local use. Spark follows it with a mixture-of-experts design and adds multimodal features to the shared chat-template lineage; both models use the same tokenizer and vocabulary.
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 use the perception encoder, while audio is converted into tokens at a rate of 12.5 per second. These details explain how the model represents varied inputs; they do not by themselves establish how accurately it interprets them.
The one-million-token limit applies to text and multimodal tokens in the same context, according to the source. Actual use of that capacity also depends on memory and inference setup. The source’s own estimate puts the key-value cache for a full one-million-token context at about 266 gigabytes in BF16, in addition to the memory needed for model weights.
large memory server for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Performance and Access Terms
The provided material does not give independent benchmark results, comparative coding scores or evidence about how reliably the model completes extended tasks. Meta’s characterization of its training and capabilities should be read as a description of the model, not proof that it outperforms alternatives.
The source also does not state the release date, the model’s license, or the full terms governing use and redistribution. It says the NVFP4 checkpoint has negligible accuracy loss, but supplies no measurements or test conditions here. Those details matter to developers assessing whether they can use the weights and what trade-offs to expect.
As an affiliate, we earn on qualifying purchases.
Deployment and Evaluation Steps
The source points developers to the Muse-Spark-Open-816B-A42B and Muse-Spark-Open-816B-A42B-NVFP4 checkpoint repositories, along with deployment recipes for Transformers, llama.cpp, vLLM and SGLang. It also describes fine-tuning guidance. Teams evaluating the release can use those materials to assess hardware needs and test the model on their own multimodal coding tasks.
Further clarity will depend on details beyond the supplied description: the applicable license, reproducible evaluations, and evidence about performance across coding and multimodal workloads. The source material does not announce a separate release date or next milestone.
As an affiliate, we earn on qualifying purchases.
Where I land
My assessment is that Muse Spark Open is a significant infrastructure-scale release: its combination of open weights, multimodal input and a coding-agent focus gives developers a model they can evaluate against real workflows. The one-million-token context and active-parameter count are useful specifications, but they do not establish accuracy or practical cost.
The strongest counterargument is that the release may be difficult to use in practice: a 1.5-terabyte BF16 checkpoint and substantial context-cache needs put deployment beyond many teams, while the supplied material offers no independent results to justify that expense. I would revise my assessment upward if reproducible evaluations showed reliable gains on long coding tasks and multimodal debugging at a clear cost advantage. Clear licensing terms and measured NVFP4 accuracy would also affect how broadly developers can adopt it.
Key Questions
What is Muse Spark Open?
It is a Meta open-weight model designed for agentic coding and multimodal input. It accepts text, images, audio and video, and generates text.
How large is the model?
The model description lists about 816 billion total parameters and roughly 42 billion active per token. All experts still need to be held in memory for deployment.
What hardware does it require?
The source says the BF16 checkpoint runs in Transformers across four nodes with eight H200 GPUs each. Its NVFP4 checkpoint is specified for one B200 node. Requirements may vary with the inference software and workload.
Does the release prove it is a leading coding model?
No independent benchmark evidence is included in the supplied material. Meta’s coding and capability descriptions are claims about the model’s design and intended use; comparative performance remains unestablished here.
Source: Hugging Face
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
