TECH Oct 17, 2026 · by Skills Team

Muse Spark Open: Meta's 816B Open Model Explained

Meta's Muse Spark Open is an 816B-parameter open-weight AI model built for coding and agentic work. Here is what it can do, what it demands, and how businesses should evaluate it.

Muse Spark Open: Meta's 816B Open Model Explained

The Short Answer

Meta has just released Muse Spark Open, an open-weight AI model it calls the most capable US open-weight model to date. Released on October 15, 2026, it puts a frontier-class coding and agentic model in the hands of any team willing to download its weights. Muse Spark Open is built for long-running agentic work, complex…

Meta has just released Muse Spark Open, an open-weight AI model it calls the most capable US open-weight model to date. Released on October 15, 2026, it puts a frontier-class coding and agentic model in the hands of any team willing to download its weights. Muse Spark Open is built for long-running agentic work, complex software engineering, and multi-step reasoning, and it can run on a desktop setup instead of only in a data center.

What Is Muse Spark Open?

Muse Spark Open is the open-weight version of Meta's Muse Spark model family. "Open-weight" means Meta publishes the trained model weights, so developers can download, inspect, fine-tune, and deploy the model on their own infrastructure. That is a sharp contrast to closed models, which you can only access through a provider's API, on the provider's terms, with your data flowing through the provider's servers.

The open release builds on Muse Spark 1.2, which shipped in August 2026, and adds extra safety post-training so the model holds up as a standalone model without the guardrails of Meta's own services. Meta says the model met the release thresholds in its Meta Superintelligence Scaling Framework, with a fuller evaluation report promised later. Meta trained the model on real developer workflows, including multi-file refactors and long debugging sessions.

It is also natively multimodal. It accepts text, images, documents, audio, and audio-video as input and returns text output. That means one model can read a screenshot of a bug, listen to a meeting recording, and review a document. Teams exploring AI and machine learning services will recognize this as the direction modern AI work is heading.

Muse Spark Open Key Specs and Capabilities

The headline numbers are striking. Muse Spark Open uses a mixture-of-experts (MoE) architecture with 816 billion total parameters, of which 42 billion are active for each token. In an MoE model, only a small subset of the network switches on for any given token, so the model holds the capacity of a very large network while the compute spent per token is closer to that of a much smaller dense model.

SpecificationMuse Spark Open
Total parameters816 billion (mixture-of-experts)
Active parameters per token42 billion
Context window1 million tokens
Input modalitiesText, image, document, audio, audio-video
Output modalityText
ArchitectureMixture-of-experts reasoning model
LanguagesTrained on data covering more than 100 languages
LicenseMuse Model License, with limited commercial terms
Weight formatsBF16 and NVFP4, plus smaller community formats

The 1 million token context window is one of the most practical features. A context window is the amount of information the model can hold in a single conversation. One million tokens is enough to feed in a large codebase, a long contract, or hours of meeting transcripts and still ask coherent questions about the whole set.

Benchmarks tell a mixed but encouraging story. In Meta's own table, Muse Spark Open scored 61.8 on the JobBench agent test, more than twice the 31.1 recorded by Inkling, the US open model from Thinking Machines Lab, while Nvidia's Nemotron 3 Ultra trailed at 27.2. It also scored 89.2 on MCP-Atlas, a test of agent and tool use, ahead of GLM 5.3 at 84.2 and Kimi K3 at 82.3. On pure coding benchmarks, Chinese open models lead: Kimi K3 scored 85 on Terminal-Bench 2.1, GLM 5.3 scored 83.9, and Qwen3.8 scored 82.0, against Muse Spark Open's 76.4. The pattern is clear: Muse Spark Open is strongest on agentic, tool-using work, and competitive rather than dominant on raw coding tests.

Meta also demonstrated real workflows rather than only benchmarks. Using Muse Code powered by Muse Spark Open, the model shipped real changes to Ice Cubes, an open-source Mastodon client, covering bug fixes, new features, and a full build. It also drove DaVinci Resolve, reviewing footage frame by frame to find the strongest moments and crop them vertically.

What Businesses Can Actually Do With Muse Spark Open

Benchmarks are useful, but the real question is what a business can do with the model on Monday morning. Here are the most practical use cases.

1. Private, on-premises coding assistance

Because the weights are open, a business can run the model inside its own environment and keep proprietary code on its own machines. That is a major advantage for teams whose code, contracts, or customer data cannot leave their control. A self-hosted coding assistant can review pull requests, explain unfamiliar code, and suggest refactors. This is the same self-hosting logic behind sovereign AI and self-hosted AI models for business: control over data and infrastructure is itself a feature.

2. Long-running agentic workflows

Muse Spark Open is built for agents that carry out multi-step tasks. A business could set it to triage a backlog of support tickets, migrate a module from one framework to another, or process a batch of documents against a checklist. Teams that already follow always-on AI agents for business will find this model fits naturally into that pattern, with the added benefit that the agent runs on infrastructure the business controls.

3. Multimodal document and media processing

Because the model accepts images, audio, video, and documents, it can process the mixed media businesses actually produce. It can summarize recorded meetings, extract structured data from scanned documents, or pull key moments from video footage, as Meta's DaVinci Resolve demonstration showed. Combined with solid data engineering and analytics practices, this turns unstructured content into usable business data.

4. Fine-tuning for a specific domain

Open weights can be fine-tuned. A business can adapt the model on its own documents, support history, product manuals, or codebase so it speaks the language of that business. Closed models generally cannot be fine-tuned to the same depth. Fine-tuning an open model keeps the adaptation in-house.

5. Prototyping AI features without vendor lock-in

Startups and product teams can prototype AI features on hosted APIs first, then move the same model in-house when scale, privacy, or unit economics justify it. Because the weights are the same, behavior stays consistent across hosted and self-hosted deployments.

The Catch: Hardware, License, and Safety Limitations

Muse Spark Open is powerful, but it is not free of friction. Any honest evaluation has to cover the limitations.

Hardware demands are enormous

Running the full model locally is a serious infrastructure project. The full-precision BF16 checkpoint is around 1.6TB in size, and the compressed NVFP4 version is around 530GB. Meta's model card says the NVFP4 build needs at least 700GB of GPU memory, and BF16 needs at least 2TB. On the server side, the BF16 checkpoint runs across four nodes of eight H200 GPUs, while the NVFP4 version serves on four Blackwell Ultra GPUs or eight H200s.

Meta's headline desktop story is that the model can run on two Apple Mac Studio computers with M5 Ultra chips, each offering up to 512GB of unified memory, linked together through Apple's MLX framework. Community quantized formats reportedly go smaller, but those trade accuracy for size. The practical takeaway: most businesses will not run the full model in-house on day one. Hosted API access through launch partners such as Together AI, Baseten, Fireworks AI, Groq, Modal, Nebius, and OpenRouter is the realistic starting point, and the weights are also available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.

The license is open, but not unconditional

Muse Spark Open is released under the Muse Model License, which includes the Muse Spark Usage Policy. The license allows use, modification, fine-tuning, and distribution, and it allows many developers to use the model without charge. But there are conditions. Products built on the model must display a "Built with Muse Spark Open" notice. The usage policy bars certain sensitive uses, including fake engagement, impersonation, and disinformation. And larger service providers above a revenue threshold must sign a separate commercial agreement before offering the model as a service.

Safety still needs human oversight

Meta added safety post-training to the open release and says the model met its internal release thresholds, but open weights mean the safety guardrails that normally sit around a hosted model are in the deployer's hands. A business running the model is responsible for its own content filters, output review, logging, and abuse controls.

Muse Spark Open vs Closed Models: How to Evaluate the Choice

The real decision for most businesses is not whether Muse Spark Open is good, but whether an open model or a closed model fits a given use case. Use these questions to decide.

Where does the data live?

If the work involves proprietary code, confidential documents, or regulated customer data, self-hosting an open model keeps that data inside your boundary. If the work uses public or low-sensitivity data, a closed API may be simpler and faster to start with.

How much control do you need?

Open weights give you control over fine-tuning, safety filters, logging, versioning, and deployment. Closed models give you convenience and managed infrastructure but leave those decisions to the provider.

What does the task actually require?

Match the model to the task. For long-horizon agentic work with heavy tool use, Muse Spark Open's benchmark strengths make it a strong candidate. For raw coding throughput, compare it against alternatives on your own code, not only on public benchmarks.

What is the total operational burden?

An open model shifts cost from API fees to infrastructure, engineering time, and maintenance. A closed model shifts it the other way. Many teams land on a hybrid pattern, using hosted APIs for prototyping and self-hosted open models for the sensitive or high-volume workloads where control matters most.

A simple evaluation checklist

  • Define 10 to 20 real tasks from your actual work, covering code, documents, and multi-step workflows.
  • Run each candidate model on those tasks and score accuracy, completeness, and output quality.
  • Measure latency and throughput under realistic load, not in a single demo run.
  • Test failure modes: what happens when the model is wrong, when tools fail, or when the input is ambiguous?
  • Confirm the license terms for your intended commercial use, and confirm your data stays where your policy requires.
  • Assign an owner for safety review, logging, and monitoring before any production rollout.

Frequently Asked Questions

What is Muse Spark Open?

Muse Spark Open is Meta's open-weight AI model, released on October 15, 2026. It has 816 billion total parameters with 42 billion active per token, a 1 million token context window, and native support for text, image, document, audio, and audio-video input with text output. Meta describes it as the most capable US open-weight model to date.

Is Muse Spark Open free to use?

The weights are available to download under the Muse Model License, and many developers can use the model without a license fee. There are conditions, including a required "Built with Muse Spark Open" notice in products built on it, usage policy restrictions, and a separate commercial agreement for larger service providers above a revenue threshold.

Can a normal business computer run Muse Spark Open?

Not the full model. The NVFP4 build needs at least 700GB of GPU memory and the BF16 build needs at least 2TB, according to Meta's model card. Meta demonstrated it running on two linked Mac Studio computers with M5 Ultra chips and 512GB of unified memory each. For most businesses, a hosted API is the practical way to start.

How is Muse Spark Open different from closed models?

With closed models, you access the model only through the provider's API, on the provider's terms, with your data flowing through the provider's servers. With Muse Spark Open, you can download the weights, fine-tune them, and run the model on your own infrastructure, keeping data in-house. The trade-off is that you take on hosting, safety controls, monitoring, and maintenance yourself.

What is Muse Spark Open best at?

Its strongest results are in agentic, tool-using work. It scored 61.8 on the JobBench agent test and 89.2 on MCP-Atlas, leading the open models in Meta's comparison table on both. On pure coding benchmarks such as Terminal-Bench 2.1, leading Chinese open models currently score higher.

Should my business switch to Muse Spark Open now?

It depends on your workload. If you need a self-hosted model for sensitive code or data, or you want to build agentic workflows you fully control, it is a strong candidate worth evaluating on your own real tasks. If you need the absolute frontier of capability, or you want to start with minimal infrastructure work, a hosted or closed option may still fit better.

Muse Spark Open is a milestone for open AI. A model of this scale, built for agentic coding work, with a million-token context window, is now available for any team to download, adapt, and run. The hardware demands are real and the license has conditions. But for businesses that value control over their data, freedom from vendor lock-in, and the ability to build agents on infrastructure they own, it is the most important open release of the year so far. If you want help evaluating Muse Spark Open for your own codebase or workflows, contact our team and we will help you plan a pilot that fits your business.

#Muse Spark Open#Meta AI#Open-Weight Models#AI Coding#Agentic AI#Machine Learning