You can host the model. You still have to build the agent.
You can put Meta’s coding model on an endpoint in your own AWS account. You cannot put an engineering process there with it.
Muse Spark Open is now listed in Amazon SageMaker JumpStart, so a team can deploy it to a real-time SageMaker AI endpoint they configure. The moment the status reads InService, that endpoint still does not hand you a browser, a set of subagents, a computer-use environment, or a sandbox that runs your tests. Amazon’s deployment post is clear on the split. Meta describes the model as trained to browse the web, coordinate subagents, and operate computers across desktop, mobile, and browser. On SageMaker, those are model capabilities. Your application, or the agent framework you already run, has to supply the tools.
That is the useful way to think about AI agents inside an engineering org. Hosted weights are not the automation. A long-running coding agent, the kind that plans a change, edits several files, watches the tests, and revises, is a product you assemble around the model. The buying questions are duller than the demo: can you hold the weights in memory, do the license terms fit what you want to ship, and does the repository stay on infrastructure you control?
What the numbers actually say
According to Meta, Muse Spark Open is an open-weight reasoning model built for coding and agentic work, including long-horizon software tasks. Meta’s published specs are 816 billion parameters, 42 billion active for each token, and a 1 million-token context window. Meta also calls it the most capable US open-weight model to date. That is a company claim. The release points at benchmarks. The deployment notes do not give you an auditable scorecard, and the listing does not settle an independent comparison.
The active-parameter figure is the one most likely to be briefed wrong. Forty-two billion active parameters are not a smaller brain that “actually thinks, ” and they do not turn this into a 42-billion-parameter model with a large closet attached. Think of sparse activation as a small reading desk in a very large library. The desk is why each token costs less compute. The library is why you still have to hold all 816 billion parameters in GPU memory. Amazon’s post treats instance choice as a capacity decision for that reason.
Meta describes the model as natively multimodal. AWS is more specific, and the difference matters. Inputs can include text, images, documents, and audio or video. Output is text only. A screenshot, a design document, or a recorded bug report can go in. Pictures and video do not come back out. The SDK identifiers in Amazon’s walkthrough use a vlm prefix. Read that as a label for those inputs, not as a second product.
The same AWS post relays Meta’s account of the training target: multi-file changes, long debugging sessions, tool calling, and a loop of planning, acting, observing, and correcting, rather than one tool call and a stop. The endpoint can return that reasoning beside the final answer. The loop still ends at the model boundary unless your agent layer keeps it going.
The license will not read itself
Open weight, here, means the weights are downloadable under the Muse Model License. That is not an unrestricted grant. It also does not mean the only way to use the model is a closed API. Meta says the license carries limited commercial terms that still let many developers use the model for free. It does not spell out the revenue or user cutoffs in that description. Do not assume an older Meta license’s numbers carried over.
Three questions belong in a legal read, not on a slide. Where does the free-use line sit? Is a customer-facing product allowed? What does the usage policy still require after you host the weights yourself? AWS points to a Muse Spark Usage Policy alongside the license and gates the listing. Studio asks you to review and accept an end-user license agreement. The SDK path requires an explicit acceptance flag. The walkthrough does not describe how that policy is enforced once the endpoint is in your account.
On safety, Meta says the model received additional safety post-training, was evaluated under its Meta Superintelligence Scaling Framework, and met that framework’s release thresholds, with a fuller report described as coming later. No scores or risk categories were published with the SageMaker announcement. Meeting a threshold you have not seen is not a finding. If this model will touch production code, the review is yours.
Size the memory, then the quality
JumpStart lists two variants, according to the AWS post. One is BF16. The other, NVFP4, stores weights in NVIDIA’s 4-bit floating-point format. Amazon says the 4-bit copy needs roughly a quarter of the memory of the BF16 weights. That is an approximate comparison in the post, not a quality study. No measured quality difference between NVFP4 and BF16 was published. A difference is possible. The listing does not tell you how large it is.
Amazon maps NVFP4 to ml.p5e.48xlarge and BF16 to ml.p6-b300.48xlarge. AWS’s EC2 documentation for the P6-B300 class specifies eight NVIDIA Blackwell Ultra GPUs and 2, 144 GB of HBM3e. It describes that class as suited to large inference jobs, including mixture-of-experts and reasoning models. That is context for the kind of machine behind the BF16 recommendation. It is not a measured footprint for this model, and the P5e mapping lives in the JumpStart post rather than on that hardware page. If the service quota is missing, the deploy will not invent it. Request it before the demo.
The 4-bit variant is the cost probe. The production choice waits on your own code. In the JumpStart announcement, the model is available in US East (N. Virginia) and US East (Ohio). Other regions are a check inside JumpStart, not a promise. AWS says a deploy typically takes 15 to 20 minutes while the instance comes up and the weights load. Ready means InService. You can pin a model version so it changes only when you redeploy. The programmatic walkthrough calls for SageMaker Python SDK v3 and an execution role that can create models and endpoints. The managed full-access policy is enough for that walkthrough. Scope the role down before anything production-shaped.
SageMaker is one hosting path, not a lock-in. Meta also places the weights on Google Cloud Model Garden, Hugging Face, and Microsoft Foundry, and offers API access through other inference providers. Its release post does not establish Amazon Bedrock as one of those paths. Meta separately claims a desktop route: the first model in its class, it says, to run on two Apple M5 Ultra Mac Studios via MLX, citing up to 512 GB of unified memory per machine and 1.2 TB/s of memory bandwidth. That is a vendor hardware claim. Its only job here is to show that “your infrastructure” is not automatically a cloud endpoint.
Private is a setting
The reason to host at all is the repository. The endpoint runs in your AWS account. Source code and prompts are processed on infrastructure you configure. You choose the instance type and the scaling limits. You can attach the endpoint to an Amazon VPC, encrypt the storage volume with an AWS KMS key, and turn network isolation on.
Do not brief the board that prompts never leave a VPC unless someone has actually attached the VPC and turned isolation on. The broader SageMaker JumpStart product page says customer data is encrypted, is not used to train the underlying models, and does not leave the customer’s VPC. That is general product language. It is not a demonstrated default for this endpoint. Network, encryption, identity, and the version pin are part of the design. They are not a follow-up ticket.
The call is easy to write. The idle box is not.
Inference, as AWS documents it, is an OpenAI-style chat request: a list of messages, an optional stream, and a token cap. Reasoning can come back in a separate field, and those tokens count against the cap. The sample in the post sets the cap at 4, 096. Leave room for the reasoning and the answer, or the ceiling can crowd out the reply you wanted.
No rate, throughput, latency, or concurrency figure ships with the announcement. Hosting instances are multi-GPU, and you pay for an endpoint for as long as it is up, whether or not anyone is calling it. The SageMaker price page is where the rate lives. A pilot nobody owns will keep that line warm.
What is Muse Spark Open, in terms a board can use without overclaiming?
Meta’s open-weight coding model, with published specs of 816 billion parameters, 42 billion active per token, and a 1 million-token context window. AWS’s JumpStart post is the guide for hosting it in your account. “Most capable” is Meta’s characterization, not a score you can audit from the listing.
Does the mixture-of-experts design make serving cheap?
It lowers compute per token. It does not remove the need to hold the full weight store in GPU memory. AWS points to NVFP4 on ml.p5e.48xlarge as the smaller-memory option and maps BF16 to ml.p6-b300.48xlarge. No quality delta between those variants was published.
Is the endpoint private by default?
It runs in your account, on infrastructure you configure. VPC attachment, KMS encryption, and network isolation are settings you choose. They are not on until you turn them on.
Can we ship this commercially, and what if we leave the endpoint up?
Not without a license read. The Muse Model License has limited commercial terms. Meta says many developers can use it free, but the thresholds are not in the public description, and SageMaker requires you to accept the license before deploy. Separately, the endpoint is billed while it runs, idle or not. Rates are on the SageMaker pricing page, not in the model listing.
What has to be true before we standardize?
The same multi-file, tool-using tasks have to have been run on your code. Hosted weights will not browse, spawn subagents, or drive a sandbox by themselves. Schedule deletion before the pilot starts, not after the invoice arrives.
Run it on your code, then turn it off
I would not standardize on either weight format, or on SageMaker versus another host, until those tasks have actually been run. First-attempt accuracy and long-horizon self-correction are descriptions Meta is offering. The test suite is the scoreboard attached to your budget.
If the pilot is real, the list is short. Read the Muse Model License and write down the answers on commercial use. Confirm the region and the GPU quota. Put the tools and the sandbox in the agent layer, not in a slide. And delete the endpoint when the week is over. An 816-billion-parameter colleague is only a bargain if it does not keep the lights on after everyone has gone home.