Amazon Web Services (AWS) has added Meta’s newly released Muse Spark Open artificial intelligence model to Amazon SageMaker JumpStart, giving enterprise cloud customers and software developers direct access to deploy the coding-focused system within their own AWS accounts. The cloud availability follows Meta’s October 15, 2026 announcement of the open-weight architecture, which the company designed for complex software engineering, multi-step reasoning, and automated agentic workflows.

Through the machine learning hub of Amazon SageMaker AI, engineering teams can provision the model to a real-time inference endpoint without manually configuring the underlying serving infrastructure. According to Meta AI Research, Muse Spark Open is a natively multimodal mixture-of-experts (MoE) model featuring 816 billion total parameters, 42 billion active parameters per token, and a context window of up to 1 million tokens.

Architecture Built for Agentic Coding and Reasoning

Meta engineered Muse Spark Open specifically around developer repository workflows, including code generation, long debugging sessions, and multi-file refactoring. Because the model supports a 1 million-token context window, enterprise engineering teams can feed entire software projects—including legacy codebases, dependency trees, and thousands of individual files—into a single prompt without splitting tasks into smaller chunks.

Unlike standard text-only generators, Muse Spark Open accepts text, images, documents, audio, and video as inputs, returning text outputs. As a reasoning model, it works through an operational plan before executing an action. When hosted on a SageMaker AI endpoint, the system returns its internal reasoning steps alongside the final output, allowing developers to audit how an automated agent arrives at a specific code modification or tool call.

While the model is trained to coordinate subagents, browse the web, and operate across desktop, mobile, and browser environments, AWS deployment rules require the customer’s application or agent framework to supply the external tools, browser interfaces, and execution environments that the model invokes.

In its release announcement, Meta chief of AI Alexandr Wang called the release “the most capable US open-weight model to date,” positioning the software against prominent open-weight releases from Chinese developers such as Alibaba’s Qwen, Moonshot AI’s Kimi K3, and Z.ai’s GLM series, according to the Guardian. However, while Meta states the model excels at long-horizon engineering tasks, reporting from Reuters notes that independent benchmark evaluations substantiating its comparative performance against rival models were not immediately available at launch.

Deploying Muse Spark Open on AWS Infrastructure

Although sparse activation keeps the per-token compute cost of Muse Spark Open closer to that of a smaller dense model, serving the architecture still requires enough GPU memory to hold all 816 billion parameters. To accommodate different hardware budgets and performance requirements, reproduced AWS deployment instructions published by Jat Ai News show that SageMaker JumpStart offers the model in two distinct variants:

  • BF16 Full-Precision Variant (Muse-Spark-Open-816B-A42B): Recommended for deployment on ml.p6-b300.48xlarge instances.
  • NVFP4 Quantized Variant (Muse-Spark-Open-816B-A42B-NVFP4): Recommended for deployment on ml.p5e.48xlarge instances, utilizing four-bit floating-point quantization to reduce memory overhead.

At launch, AWS availability for these deployments is listed in the US East (N. Virginia) and US East (Ohio) Regions. Developers can initiate deployments either through the visual interface in Amazon SageMaker Studio or programmatically via the SageMaker Python SDK.

Access to the model weights on JumpStart is gated. Before spinning up an endpoint, AWS users must review and accept an end-user license agreement governed by the Muse Model License and the Muse Spark Usage Policy. Enterprise administrators can also configure optional security controls during deployment, including Amazon Virtual Private Cloud (VPC) routing, AWS Key Management Service (KMS) encryption keys, and strict network isolation.

Broader Ecosystem and Local Hardware Support

Beyond Amazon SageMaker AI, Meta is distributing the model weights across Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. For teams preferring managed serverless inference over self-hosted cloud instances, third-party API access is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

The model is compatible with major open-source inference engines, including vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, Unsloth, and SGLang. Within SGLang, Meta included custom auto-tunable Helion GPU kernels that deliver 1.03x to 3.95x kernel-level speedups across NVIDIA and AMD GPUs, alongside a 3% end-to-end speedup on AMD MI350 hardware.

In an unusual hardware pairing for an 816-billion-parameter system, Meta also optimized Muse Spark Open for local desktop execution on Apple silicon using the MLX framework. By linking two Mac Studios equipped with Apple M5 Ultra chips—each supporting up to 512 GB of unified memory and 1.2 TB/s of memory bandwidth—developers can pool a full terabyte of memory to run the entire model locally without cloud infrastructure.

Frequently Asked Questions

What AWS instance quotas are required before deploying Muse Spark Open?

Because the 816-billion-parameter model requires high-capacity GPU clusters, customers must verify that their AWS account holds sufficient SageMaker endpoint service quotas for either the ml.p5e.48xlarge instance (for the NVFP4 quantized model) or the ml.p6-b300.48xlarge instance (for the BF16 variant) in the US East (N. Virginia) or US East (Ohio) Regions before attempting deployment.

How should enterprises choose between the BF16 and NVFP4 variants?

The choice depends on hardware availability and workload sensitivity. The BF16 variant preserves full 16-bit brain floating-point precision on newer ml.p6-b300.48xlarge hardware, while the NVFP4 variant compresses the weights to fit onto ml.p5e.48xlarge instances. AWS and Meta documentation have not yet published separate comparative figures for endpoint pricing, deployment spin-up times, or token throughput differences between the two configurations.

What licensing terms apply to Muse Spark Open on SageMaker?

Unlike closed proprietary models from OpenAI or Anthropic, Muse Spark Open allows developers to download and modify the underlying parameters via PyTorch. However, it is released under the custom Muse Model License, which enforces the Muse Spark Usage Policy. On AWS, users must formally accept the gated JumpStart End-User License Agreement (EULA) prior to provisioning an endpoint.

Sources

Read More