AWS says Meta Muse Spark Open is available in Amazon SageMaker JumpStart, the model hub of SageMaker AI. The coding-optimized, open-weight model launches in US East (N. Virginia) and US East (Ohio). Users can deploy it through SageMaker Studio or the SageMaker Python SDK to a real-time endpoint in their own AWS account.

Context for substantial coding tasks

Muse Spark Open has a one-million-token context window. Training included code generation, debugging, multi-file refactoring and long-running tasks in real repositories. Meta emphasizes agentic coding, first-attempt accuracy, tool calling and understanding large codebases. These are stated areas of focus, rather than independently established performance comparisons.

The reasoning model accepts text, images, documents, audio and audio-video input, and produces text. AWS identifies assistance with private repositories, long-running coding agents and turning screenshots, design documents or recorded bug reports into reproduction plans or draft fixes as potential applications.

According to Meta, the model plans, acts, observes results and corrects itself during extended tasks. Training also covered web browsing, subagent coordination and computer operation across desktop, mobile and browser environments. On SageMaker AI, the application or agent framework must supply the tools, browser and computer-use environment.

Two variants, with different memory requirements

The mixture-of-experts architecture contains 816 billion parameters, with 42 billion active per token. Sparse activation reduces computation per token, but serving still requires enough GPU memory to hold all parameters. The active parameter count therefore does not describe the full memory requirement.

• BF16: Muse-Spark-Open-816B-A42B runs on ml.p6-b300.48xlarge.
• NVFP4: Muse-Spark-Open-816B-A42B-NVFP4 runs on ml.p5e.48xlarge.

NVFP4 uses NVIDIA's four-bit floating-point format and approximately one-quarter of BF16's weight memory. AWS recommends evaluating both variants on the intended workloads before choosing one for production. The memory figure alone does not establish equal quality or a particular reduction in costs.

Deployment prerequisites and controls

Deployment requires an AWS account with permission to create models and endpoints, a SageMaker AI IAM execution role, sufficient quota for the selected endpoint instance type and a supported region. Studio also requires a SageMaker AI domain with Studio access. The SDK route requires Python 3.10 or later and SageMaker Python SDK v3. AWS says AmazonSageMakerFullAccess is sufficient for its walkthrough, while recommending production roles restricted to necessary actions and resources.

The weights are released under the Muse Model License, including the Muse Spark Usage Policy. Access is gated: users must accept the EULA in Studio or set accept_eula=True in an SDK deployment. Availability of the weights therefore comes with license conditions.

SageMaker AI handles the serving infrastructure. Deployment controls include VPC attachment, storage-volume encryption with an AWS KMS key, container network isolation, instance selection, scaling limits and model-version pinning.

From deployment to a response

In Studio, the walkthrough directs users to find the model in JumpStart, select a variant and choose Deploy. They then configure the endpoint name, a supported instance type and an initial instance count of one. AWS says deployment typically takes 15–20 minutes; the endpoint is ready at InService. The SDK can use the Studio domain's execution role, or role_arn in another Python environment with AWS credentials. The model_version setting in JumpStartConfig pins a version.

The endpoint accepts an OpenAI-compatible chat payload with a messages list. The answer appears in message.content, reasoning in message.reasoning_content and reasoning-token usage in usage.reasoning_tokens. AWS advises budgeting max_tokens for both reasoning and the answer. Streaming requires stream=true and endpoint.invoke_with_response_stream instead of invoke.

Hosting charges continue while idle

The hosting instances use multiple GPUs. Endpoint charges accrue while an endpoint runs, even without requests. AWS instructs users to delete unused endpoints to stop those charges. SDK cleanup covers the model, endpoint and endpoint configuration; Studio deployments can be removed through the Endpoints page in the SageMaker AI console.