Muse Spark Open: Meta’s 816B Model and the Licence Line You Must Read

Muse Spark Open is Meta's biggest open-weight release: 816B parameters, 1M context and a licence with strings attached. What it means for builders.

Disclosure: this article is based on Meta’s release post, the model card as reported by several outlets and third-party coverage. We have not self-hosted Muse Spark Open, which needs data-centre class hardware. Figures are attributed in the text.

Meta released Muse Spark Open on October 15, 2026, and called it the most capable US open-weight model to date. It is an 816-billion-parameter mixture-of-experts model that activates 42 billion parameters per token, with a one-million-token context window and native multimodal input. The weights are on Hugging Face and the major clouds.

The honest summary for builders: Muse Spark Open is a serious model with a real catch. It trails the best Chinese open models on Meta’s own coding table, and its licence requires any product built on it, or on anything distilled from it, to prominently display a credit in the user interface. If you can live with that line, it is worth testing. If you cannot, the MIT and Apache licensed alternatives are strong enough that you will not miss much.

Table of Contents

Muse Spark Open library shelves full of books
Open weights are not the same as an open licence. Muse Spark Open lets you download everything and still tells you how to label your product.

What Muse Spark Open Is

Muse Spark Open is the open-weight version of Meta’s Muse Spark model. According to Meta’s post on the PyTorch blog, it shares the same base model as Muse Spark 1.3 and received additional safety post-training so that it holds up as a standalone model without the system guardrails Meta runs on its own hosted version.

The architecture is a mixture of experts: 816 billion parameters in total, 42 billion active per token. That ratio is what makes a model this large practical to serve. Each token pays the compute cost of a 42B model, while the full parameter set holds far more knowledge.

It is built for coding and agent work. Meta describes it as strong at code generation, debugging and codebase understanding, and pitches it at long-horizon tasks and multi-step reasoning. It accepts images as well as text.

Distribution is broad. The weights are on Hugging Face at meta-models/Muse-Spark-Open-816B-A42B, and the model is available in Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. Meta also shipped custom auto-tuned GPU kernels in SGLang, which it says deliver kernel-level speedups of 1.03 to 3.95 times on NVIDIA and AMD GPUs, and a 3% end-to-end speedup on AMD’s MI350.

The model sits below Meta’s closed frontier models. Muse Wonder and Muse Wonder Full, released three days earlier, remain closed, and Implicator’s coverage was explicit that Muse Spark Open is less advanced than Wonder.

The Muse Spark Open Licence, Read Carefully

This is the part to read before you download anything. Muse Spark Open ships under the Muse Model License, not a standard open-source licence. No Code MBA’s release tracker describes it as having limited commercial terms.

The clause that stands out, reported by The New Stack, requires any company using the model to prominently display “Built with Muse Spark Open” in the user interface of each product or service that uses it. The requirement also covers derivative models, including ones that were trained, fine-tuned or distilled from it.

Think about what that means in practice. If you fine-tune a small model on outputs from Muse Spark Open and ship it inside an app, the app must carry Meta’s credit. For a consumer product with its own brand, that may be unacceptable. For an internal tool, it may not matter at all.

Compare the alternatives. DeepSeek V4.1 Flash uses MIT. Gemma 4 and Meta’s own Muse Glimmer use Apache 2.0. NVIDIA’s Nemotron 3 Ultra uses OpenMDW-1.1, which grants royalty-free commercial use and carves model outputs out of the licence. Mistral Large 4 uses a modified MIT licence with a clause about model-as-a-service use. Muse Spark Open’s attribution requirement is the most visible obligation in this group.

The New Stack summed it up as open weights with a semi-open licence. That is fair. You get the full model to run, inspect and adapt. You do not get to forget where it came from.

How Muse Spark Open Scores

Meta’s own benchmark table is where the most telling number appears. On Terminal-Bench 2.1, a coding test, Muse Spark Open scores 76.4. Moonshot’s Kimi K3 scores 85.0 on the same table. Implicator’s headline summarised it as an 816B model that trails Chinese rivals at coding.

The New Stack’s framing was more generous: it beats US rivals and matches Chinese models on agentic tasks. Both can be true. Among American open-weight models, it is clearly the most capable. Against the best Chinese open models, it is competitive on agentic work and behind on some coding tests.

Meta says detailed results are in the model card, with a full evaluation report to follow. At release, that report was listed as coming soon. Until it arrives, independent testing is thin, and the claims rest largely on Meta’s own table.

On safety, Meta says the model was evaluated across content safety, agentic risk, privacy and preparedness under its Meta Superintelligence Scaling Framework, and met the thresholds that framework requires for release.

Running Muse Spark Open: Hardware Reality

An 816B model is not a laptop model. The rough arithmetic: at 4-bit precision, a model needs about half a byte per parameter, so the weights alone come to roughly 400 GB before any context cache. At 8-bit, double that. This is multi-GPU server territory.

The one surprise is Apple. Meta says Muse Spark Open runs on Apple M5 Ultra Mac Studios, and Tech Insider, citing Engadget, reported that Apple worked with Meta to get it running through the MLX framework. That makes it one of the very few frontier-scale open models with a desktop path, though you should expect it to need the highest memory configuration and to run far slower than a GPU cluster.

For most teams, the practical options are a managed endpoint on SageMaker, Model Garden or Foundry, or a multi-GPU server running SGLang with Meta’s tuned kernels. The managed route removes the hardware question; the self-hosted route gives you data control and, at high volume, lower cost per token.

No Code MBA’s tracker made the obvious point plainly: the model needs very large memory to self-host. If you were hoping to run Meta’s best open model on a single consumer GPU, the model you want is Muse Glimmer, not this one.

Muse Spark Open vs Chinese and European Rivals

Model Size Licence Notable
Muse Spark Open 816B, 42B active Muse Model License, attribution required Most capable US open model, 1M context
Kimi K3 Large MoE Open weights 85.0 on Terminal-Bench 2.1 in Meta’s table
DeepSeek V4.1 Flash 552B, about 8B active MIT Very cheap API, 1M context
Nemotron 3 Ultra 550B, 55B active OpenMDW-1.1 Pre-quantised NVFP4, 1M context
Mistral Large 4 About 1T, 49B active Modified MIT Strong coding and cyber focus

Khaleej Times noted that Alibaba’s Qwen, Moonshot’s Kimi K3 and Z.ai’s GLM models have been considered the global open-weight frontrunners, and that Meta is now claiming to compete on a similar level. That is the right way to read the release. Muse Spark Open does not take the lead. It brings a US-developed option into the top group.

For some buyers, that origin matters more than a few benchmark points. Government contractors and companies with policies against Chinese-developed models now have a large, capable US option that is not a closed API. National CIO Review framed it exactly that way: another US-developed open-weight choice for enterprises deciding where to run AI.

Why Meta Released It

Meta’s open-model story has twisted this year. After the Llama era, Meta retooled its AI work under Alexandr Wang and launched Muse Spark as a closed model. Then it released Muse Glimmer in August under Apache 2.0 and promised open Spark weights. Muse Spark Open is the delivery on that promise, about two months later.

Mark Zuckerberg has framed this as part of a broader argument that the US and its allies should lead the open-source AI ecosystem. Meta’s release post says open-weight models are part of how it brings the benefits of what it calls personal superintelligence to everyone, and part of how safety work advances across the field.

There is a business logic too. Open models build an ecosystem of tools, fine-tunes and developers around Meta’s architecture, which helps recruiting, helps Meta’s own products and puts pressure on rivals who only sell closed APIs. Keeping the frontier closed with Muse Wonder while opening the tier below it is a strategy several labs now follow.

Sensible Uses

Internal coding and agent platforms

Inside a company, the attribution clause is easy to satisfy. A self-hosted Muse Spark Open behind your firewall gives you a capable coding model with no data leaving your network.

Regulated or sovereign deployments

Organisations that cannot send data to third-party APIs and prefer a US-developed model have few alternatives at this capability level.

Research and evaluation

Having the full weights of a frontier-scale model lets researchers study behaviour, test safety interventions and benchmark in ways a closed API never allows.

Long-context document work

The one-million-token window and multimodal input make it a candidate for large document and codebase analysis on private infrastructure.

Self-Hosting vs a Managed Endpoint: How to Think About Cost

For a model this size, the decision between running it yourself and using a cloud endpoint is mostly arithmetic, and the arithmetic depends on how busy the model will be.

A self-hosted deployment has a fixed cost. You pay for the servers, or for reserved GPU capacity, whether the model handles one request an hour or a thousand a minute. A managed endpoint has a variable cost: you pay per token or per hour of use, and the provider absorbs idle time.

At low or spiky volume, managed almost always wins. A multi-GPU node sitting idle overnight is expensive, and the engineering time to keep it patched, monitored and scaled is easy to underestimate. At steady high volume, self-hosting can become cheaper per token, especially with the mixture-of-experts design, because only a fraction of the parameters run on each token and a well-tuned server can batch many requests together.

The break-even point is specific to your traffic, so measure rather than guess. Run a managed endpoint for a month, log the tokens processed and the hours the model was actually busy, then price the equivalent dedicated hardware. If the dedicated cost is well below the managed bill and your volume is stable, self-hosting is worth the operational overhead. If the numbers are close, the managed route buys you a lot of simplicity for a small premium.

There are non-financial reasons too. Data residency, audit requirements and the ability to pin an exact model version can justify self-hosting even when it costs more. Equally, a small team without infrastructure experience should be honest about the hidden cost of running large GPU servers. Downtime at three in the morning is a real price.

Whichever you choose, keep the application talking to the model through a standard API. Inference servers that expose an OpenAI-compatible interface make it possible to move from a managed endpoint to self-hosted hardware, or between models, without rewriting your code.

A Licence Checklist Before You Build

Licences for open-weight models have become varied enough that a short checklist saves arguments later. Run through it with whoever signs off on your software dependencies.

  • Attribution: does the licence require a visible credit in the product? Where, and in what form?
  • Derivatives: do obligations carry over to fine-tunes and to smaller models distilled from outputs?
  • Outputs: are model outputs covered by the licence, or explicitly carved out?
  • Service use: is offering the model as a hosted service to third parties restricted?
  • Thresholds: do terms change above a certain number of users or a revenue level?
  • Acceptable use: is there a usage policy attached, and does your use case sit comfortably inside it?

Write the answers down for each model you consider. MIT and Apache 2.0 answer most of these questions in your favour by default. Custom licences need reading line by line, and the answer to the derivatives question is the one most likely to surprise you, because it can follow your work long after you stop using the original model.

Who Should Use What

If you are Start with Why Link
Building internal tools on US-developed open weights Muse Spark Open Most capable US open model Get the weights
Shipping a branded consumer product DeepSeek V4.1 Flash or Gemma 4 MIT or Apache, no attribution clause See specs
Wanting the best open coding score Kimi K3 Leads Meta’s own coding table Read comparison
Running a local agent on one GPU Muse Glimmer 30B, Apache 2.0 See Glimmer

Related on this site: for more context, see Muse Wonder, Mistral Large 4 and Nemotron 3 Ultra.

Frequently Asked Questions About Muse Spark Open

What is Muse Spark Open?
An open-weight mixture-of-experts model from Meta with 816 billion total and 42 billion active parameters, a one-million-token context and native image input. It shares a base model with Muse Spark 1.3.

When was Muse Spark Open released?
October 15, 2026, three days after Meta’s closed Muse Wonder models.

Is Muse Spark Open open source?
The weights are open, but the Muse Model License has limited commercial terms and requires products using the model or its derivatives to display a “Built with Muse Spark Open” credit.

Can I run it at home?
Realistically, no. The weights are roughly 400 GB at 4-bit. Meta says it runs on Apple M5 Ultra Mac Studios via MLX, but most teams will use a managed cloud endpoint or a multi-GPU server.

How does it compare with Kimi K3?
On Meta’s own Terminal-Bench 2.1 table, Muse Spark Open scores 76.4 against Kimi K3’s 85.0. Meta says it matches Chinese rivals on agentic tasks.

Where can I get the weights?
Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry.

Verdict

If you need a large, capable, US-developed open-weight model for internal coding and agent work, Muse Spark Open is now the strongest option, and the cloud availability makes it easy to try.

The honest caveat is twofold. It trails the best Chinese open models on Meta’s own coding benchmark, and its licence attaches a visible attribution requirement to every product and derivative. Independent evaluation is still limited, and Meta’s full report had not been published at release.

Next step: read the Muse Model License with your legal team first, then spin up a managed endpoint and run your own coding tasks against DeepSeek V4.1 Flash or Kimi K3 before committing infrastructure.

Sources and Further Reading

On this site: compare self-hosting stacks in LLM inference servers compared, pick a smaller option from best local LLMs, see the wider market in AI models compared, and ground outputs with retrieval-augmented generation.