Released on October 15, 2026, Muse Spark Open is Meta’s new open-weight multimodal large model. Meta describes it as the most capable open-weight model originating from a US lab. The architecture features a total of 81 billion parameters, with roughly 42 billion active parameters activated per token. It supports up to a 1M-token context window and is built specifically for long-running agent workloads. The model accepts text, image, video and audio inputs, with design priorities focused on code comprehension, tool calling and multi-step reasoning.
Meta’s “most capable” label is bounded in scope. This ranking applies only to open-weight models developed within the United States; it is not an unconditional global benchmark. The official Meta Research release explicitly uses the phrasing “the most capable US open-weight model to date”, and the developer model page further clarifies it as “the most capable open-weight model from a US lab”. This nuance matters for developers comparing cross-border open model options.
The model weights are distributed through Hugging Face. In addition, Meta has synced Muse Spark Open to major cloud model marketplaces, including Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. For local inference and deployment, the official supported runtime stack includes SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth. These frameworks cover cloud GPU serving and Apple Silicon native execution, giving developers multiple routes to test and run the model.
At its core, Muse Spark Open is a native multimodal sparse Mixture-of-Experts (MoE) model. It can ingest four types of input modalities: text, static images, video frames and audio streams. The 1M-token long context window is the foundation for its target use case: long-horizon AI agents. For these agent workflows, the model is designed to parse lengthy codebases, execute sequential tool invocations and complete multi-stage reasoning tasks. It is critical to note that releasing open weights does not automatically grant the model permission to perform web search, execute arbitrary code, or access local desktop controls. Those capabilities must be implemented and secured by the external agent application layer built on top of the model.
Understanding 81B Total Parameters and 42B Active Parameters
The two headline numbers, 81 billion total parameters and 42 billion activated per token, are often misinterpreted. They cannot be directly converted into a simple dense GPU memory requirement. Muse Spark Open uses a sparse MoE design. Every input token routes to only a subset of the expert parameters, so only 42B parameters activate for each token even though the full model stores 81B total parameters.
Meta provides two primary weight formats for download: BF16 and NVFP4. The file footprint for BF16 weights is approximately 1.65 TB, while NVFP4 compressed weights occupy around 537 GB. These storage figures represent only the raw model weights. They do not include memory overhead from KV cache, activation tensors and runtime libraries. Real-world GPU memory consumption depends heavily on context window length, batch size, inference framework and parallelization strategy.
The “42B active” metric describes computational cost for a single token forward pass. Developers should not estimate memory requirements by treating it like a standalone dense 42B model. KV cache grows linearly with context length, which adds significant memory pressure for long 1M-token prompts. Before production deployment, teams must run load tests with target context lengths and batch configurations.
Meta’s Desktop Deployment Demo: Running on Apple Mac Studio Hardware
Meta showcased local desktop inference using a pair of M5 Ultra Mac Studio machines for Muse Spark Open. Each M5 Ultra provides up to 512 GB of unified memory, combining two units delivers a total of 1 TB of shared memory. This hardware configuration is sufficient to load the NVFP4 quantized weights, relying on MLX and high-speed interconnects between the two Apple Silicon hosts.
This demonstration is not a guarantee that any consumer desktop can run the full 81B model. It is a defined hardware reference configuration. Meta also lists an alternative setup using four Mac Studio devices each with 256 GB unified memory. Developers must select hardware matching their weight format, target context length and parallelism scheme.
For NVIDIA hardware stacks, Meta provides reference recipes. The NVFP4 weight variant can run starting from 4 B300 GPUs using TP4 / EP4 tensor-parallel and expert-parallel splitting. The larger BF16 weight version requires 8 B300 GPUs configured as TP8 / EP8. The SGLang documentation also includes tuning guidance for NVFP4 with Flash2 prefetch and HiCache optimizations. These are example deployment templates rather than minimum hardware requirements. Actual resource consumption varies by workload.
Long-Horizon Agent Capabilities: More Than Just a Large Context Window
The core use case for Muse Spark Open is long-running agent workflows. The model is built to read code, plan modifications, invoke external tools, validate return results and iteratively adjust its strategy when tasks fail. Meta positions the model for programming agents and extended multi-turn agent sessions. Its multimodal perception stack supports end-to-end agent pipelines that combine visual media, audio and text.
The bundled capability suite includes cross-file code editing, video and image comprehension, tool invocation, multi-agent orchestration, and persistent reasoning over extensive historical context. Still, the underlying model remains stateless in terms of system permissions. The host application defines tool schemas, access permissions, result validation logic and termination rules. Downloading the open weights will not grant the model native access to the local filesystem or external APIs.
Meta Research published practical demos built on Apple Silicon hardware using Muse Code, a related capability stack. The examples include modifying open-source iOS projects and reviewing media assets inside DaVinci Resolve to generate edit decision lists. These are showcase demonstrations and do not guarantee that all code repositories or desktop software can be reliably controlled by the agent.
Choosing Weight Formats and Parallel Topologies
The following table summarizes Meta’s reference deployment configurations for each weight variant. Memory footprint and hardware requirements differ substantially, so weight selection is the first major decision in any rollout.
| Weight Format | Model File Size | Reference GPU / Hardware Setup | Parallelism Mode |
|---|---|---|---|
| BF16 | ~1.65 TB | 8 × B300 | TP8 / EP8 |
| NVFP4 | ~537 GB | 4 × B300 | TP4 / EP4 |
| Apple MLX | Unified memory loading | 2 × 512GB Mac Studio or 4 × 256GB Mac Studio | Thunderbolt 5 interconnect |
The table highlights a key rule: model weight size is not equivalent to minimum runtime memory. The KV cache and activation buffers add overhead. Meta recommends developers shrink the context length from the maximum 1,048,576 tokens down to the business requirement value, then run stress tests before production deployment. This reduces memory pressure and stabilizes throughput.
Evaluation Benchmarks: Look Beyond Single Aggregate Scores
Meta’s benchmark suite covers agent performance, knowledge and reasoning, code tasks, and multimodal tool-use scenarios. All results are generated with the xhigh reasoning preset. The published benchmark table includes DeepSWE, Terminal-Bench, JobBench, MCP-Atlas and GDPval-AA v2.
These benchmark scores are not universal performance guarantees for every real-world task. Meta’s release materials emphasize that the model targets long-duration tasks, complex software engineering and multi-step reasoning. When evaluating Muse Spark Open, engineering teams should design custom evaluation tracks that match their actual workload. Important metrics to log include task completion rate, tool call accuracy, latency and output quality across fixed datasets, tool environments and context window sizes. Benchmark numbers only serve as preliminary screening.
How Developers Can Start Testing Without 81B-Class Hardware
Teams without sufficient high-end GPU memory can first validate task behavior using hosted model endpoints before downloading the full weight files. Meta’s list of launch partners includes Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Each provider maintains its own model ID, maximum context limits and pricing rules, so developers must refer to each platform’s documentation for details.
If local weight download is required, Meta recommends installing the Hugging Face CLI first. Users select either BF16 or NVFP4 repository matching their hardware. The simplified download workflow requires completing Meta’s model license application and preparing enough disk storage.
The downloaded folder includes the tokenizer, chat template, config.json and processor assets. Developers should validate the tokenizer with the official Transformers library first, then select an inference runtime such as SGLang, vLLM or MLX. This article does not include local testing or GPU inference runs on the full weight set.
Open Weights Do Not Equal Unrestricted Commercial Use
Muse Spark Open is released under the Meta Muse Model License. The license defines redistribution terms, usage policies and specific commercial clauses for Model-as-a-Service deployments. Model metadata will tag releases with muse-model-license. Users must accept the license form, and the archive includes LICENSE.md and USE_POLICY.md documents.
The license permits use, copying, modification, merging and distribution, but redistributed copies must retain the original license, copyright and usage notices. Products or services built with the model must visibly display attribution: “Built with Muse Spark Open”.
There is a major commercial threshold for MaaS operators. If a business offers hosted MaaS using this model and records more than 20 million USD in annual revenue for 12 consecutive months, they are required to notify Meta and sign a separate written agreement before serving the model or derived variants to end users. All end-user deployments must comply with the usage policy.
This licensing framework sets clear boundaries. Developers gain full access to weights and runtime support, but production rollout requires license review and security planning.
Operational Workflow Considerations for Multi-Model Agent Systems
Teams building agent pipelines that integrate Muse Spark Open with other foundation models must manage API authentication, request throttling and routing between multiple endpoints. Many production stacks use an API gateway to consolidate traffic control. 4sapi, an API gateway, helps teams centralize credential management, enforce rate limits and aggregate logs across mixed model services in agent workflows.
When building long-context agent systems, request routing and observability become critical. Long inference jobs consume more compute and can experience variable latency. A unified gateway layer abstracts the underlying model endpoints and provides consistent access controls regardless of whether the workload runs on local hardware or cloud hosted model providers.
For teams that combine local self-hosted Muse Spark Open with other remote multimodal APIs, separating backend model serving from client application code reduces integration complexity. Gateway routing logic can route lighter tasks to smaller models while sending complex long-context agent jobs to the 81B Muse Spark Open instance. This layered architecture balances cost, latency and resource utilization.
Conclusion
Muse Spark Open represents a significant milestone for open-weight multimodal agent models. With 81B total parameters, 42B active per token, native multimodal support and a 1M-token context window, Meta targets the model for long-running agent and code engineering workloads. The sparse MoE design brings unique memory and parallelization challenges, with two weight formats offering tradeoffs between precision and storage footprint.
The release strategy balances cloud hosted access for early prototyping and open weights for self-hosted deployment, supported across NVIDIA GPU hardware and Apple Silicon Mac Studio devices. The license terms allow broad development and distribution while imposing commercial obligations for high-revenue MaaS services. Developers must treat benchmark scores only as reference, and validate model behavior on their own agent task suites.
For teams orchestrating multi-model agent pipelines, centralized API traffic governance simplifies operations. An API gateway streamlines access to a mix of self-hosted open models and third-party hosted endpoints. 4sapi provides that unified control plane for developers building complex multimodal agent systems.
International access: https://4sapi.com
Domestic access: https://4sapi.org
