# Muse Spark Open

> Source: https://aiwiki.ai/wiki/muse_spark_open
> Updated: 2026-10-15
> Categories: AI Models, Large Language Models, Meta AI, Multimodal AI, Reasoning Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Muse Spark Open." aiwiki.ai, 15 Oct 2026. https://aiwiki.ai/wiki/muse_spark_open
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Muse Spark Open** is an [open-weight](https://aiwiki.ai/wiki/open_weights) multimodal reasoning model released by [Meta Superintelligence Labs](https://aiwiki.ai/wiki/meta_superintelligence_labs) on October 15, 2026. Meta specifies 816 billion total parameters, 42 billion active parameters per token, and a 1 million-token context window.[1] It accepts text, images, video, and audio and produces text, with training focused on coding and extended agent workflows.[7]

The release provides downloadable checkpoints rather than only a hosted service. It belongs to the [Muse Spark](https://aiwiki.ai/wiki/muse_spark) family and has its own model card, evaluation results, and license.[1]

## Model overview

Meta's model card specifies:[2]

| Attribute | Specification |
| --- | --- |
| Release date | October 15, 2026 |
| Parameters | 816B total, 42B active |
| Context length | 1M tokens at inference |
| Modalities | Text/image/video/audio input; text output |
| Decoder | 62 layers, hidden dimension 8192 |
| Experts | 256 per layer, 8 selected per token |
| Attention heads | 128 query, 16 key-value |
| License | Muse Model License |

## Architecture

Muse Spark Open uses a sparse [mixture of experts](https://aiwiki.ai/wiki/mixture_of_experts). A learned projection compresses each token's representation before the selected experts process it; a second shared projection restores the original width. The router still receives the full-width representation. Baseten describes this arrangement as reducing expert parameters and communication between GPUs compared with an uncompressed design.[4]

The decoder interleaves 46 local-attention layers with 16 full-attention layers. Local layers use a 2048-token sliding window; full-attention layers retain access to the complete sequence. Baseten reports a reduction in key-value cache memory from 31 GiB to 8.4 GiB per sequence at 128K context relative to an all-global-attention configuration. That is a configuration-specific cache comparison, not a reduction of the model's entire memory requirement.[4]

The [vLLM](https://aiwiki.ai/wiki/vllm) implementation uses a [Perception Encoder](https://aiwiki.ai/wiki/perception_encoder) for images and sampled video frames, together with a separate audio encoder. Its audio path takes 24 kHz waveforms. The vLLM team states that a video's audio track is not processed automatically: applications must supply video and extracted audio as separate input parts when both are needed.[5]

## Checkpoints and local deployment

Meta distributes two reference checkpoint formats through [Hugging Face](https://aiwiki.ai/wiki/hugging_face). Its download guide gives the following approximate weight footprints and example serving configurations.[6]

| Checkpoint | Repository | Approximate weights | Example configuration |
| --- | --- | ---: | --- |
| BF16 | `meta-models/Muse-Spark-Open-816B-A42B` | 1.65 TB | 8 B300 GPUs, TP8/EP8 |
| NVFP4 | `meta-models/Muse-Spark-Open-816B-A42B-NVFP4` | 537 GB | 4 B300 GPUs, TP4/EP4 |

NVFP4 stores most weights at 4-bit precision while retaining some layers in BF16. These figures exclude the key-value cache, activations, and runtime overhead. Hardware requirements also depend on the context length, interconnect, parallelism, and serving software. Active parameter count is not the complete weight footprint.[6]

Meta's developer launch guide lists [SGLang](https://aiwiki.ai/wiki/sglang), vLLM, TensorRT-LLM, Transformers, [llama.cpp](https://aiwiki.ai/wiki/llama_cpp), [MLX](https://aiwiki.ai/wiki/mlx), and [Unsloth](https://aiwiki.ai/wiki/unsloth) integrations.[7]

Its Mac recipe requires two M5 Ultra Mac Studios with 512 GB of unified memory each, or four machines with 256 GB each, connected directly with Thunderbolt 5 and RDMA. It also specifies macOS 26.2 or later, MLX 0.30.1 or later, and disk space for the full NVFP4 checkpoint on every machine. This is a multi-machine setup, not an ordinary laptop deployment.[13]

Launch-day support can require release-specific software. The vLLM team supplied dedicated NVIDIA and AMD container images and said its implementation would reach nightly Python wheels after the corresponding code review and merge. Installing an older standard package is therefore not equivalent to using the launch integration.[5]

### Hosted access

Launch partners include Baseten, Fireworks AI, Groq (limited preview), Modal, Nebius, OpenRouter, and Together AI. Meta also lists Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models as distribution catalogs.[7]

## Prompting and tool use

Muse Spark Open always reasons before answering. Meta's prompting guide exposes four settings through `reasoning_effort`; `xhigh` is the default. Values such as `none` and `max` are rejected.[8]

| Reasoning effort | Reasoning strength |
| --- | ---: |
| `low` | 8 |
| `medium` | 32 |
| `high` | 256 |
| `xhigh` | 512 |

An explicit numeric `Reasoning strength` line in the system prompt takes precedence over the request parameter. Servers configured with the `muse` reasoning parser separate reasoning from the final answer. Meta recommends streaming at higher strengths to avoid timeouts.[8]

For tools, the application supplies function definitions, executes requested calls, and returns their results as tool messages. The model can request multiple calls in one response, but the surrounding application supplies the permissions and execution loop. Model tool-calling capability does not itself grant access to a computer or external service.[8]

## Evaluation

Meta's release table mixes internal runs, benchmark-provider results, and other vendors' self-reports, using `xhigh` for Muse Spark Open. Meta cautions that its setup may not be optimal for competing models.[9]

Selected release-card results:[2][9]

| Benchmark | Muse Spark Open | Configuration or interpretation |
| --- | ---: | --- |
| Artificial Analysis Intelligence Index | 39 | Version 4.3, source snapshot dated October 7, 2026 |
| OSWorld Verified | 80.4 | Computer-use evaluation |
| HLE-Diamond | 26.1 | Without tools |
| DeepSWE 1.1 | 60 | Muse Code harness |
| Terminal-Bench 2.1 | 76.4 | Terminus 2 agent |
| Video-MME v2 | 61.5 | Subtitles included, no audio |

Meta attributes DeepSWE to [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) runs with [Muse Code](https://aiwiki.ai/wiki/muse_code), while Terminal-Bench uses Terminus 2.[9]

[OSWorld](https://aiwiki.ai/wiki/osworld) checks task completion in real computer environments.[10] Meta used a 361-task Ubuntu split with GUI-only tools, a 200-step limit, and four attempts per task, reporting mean reward.[9]

## Customization and inference optimization

The release includes training integrations as well as inference checkpoints. Meta's developer guide lists supervised fine-tuning, low-rank adaptation, and reinforcement-learning paths through TorchTitan, SGLang with Miles, NVIDIA NeMo AutoModel, Hugging Face tools, and managed training APIs. Developers still need task-specific data, evaluation, and infrastructure for the selected method.[7]

The PyTorch team released Helion implementations of operations including RoPE, RMSNorm, and quantization. It reported 1.03-3.95x speedups for individual kernels and a 3% end-to-end throughput improvement on AMD MI350. Kernel speedups and complete-system throughput measure different things; the larger range is not a claim of a fourfold speedup for every inference workload.[11]

[Speculative decoding](https://aiwiki.ai/wiki/speculative_decoding) is another supported path. The vLLM integration uses the separately trained DSpark draft model to propose up to seven tokens before verification by Muse Spark Open.[5]

## License and access conditions

The Muse Model License grants rights to use, modify, fine-tune, distill, merge, distribute, and sell the software subject to its conditions. Copies and derivatives must retain the copyright notice, license, and usage policy. Products using the model or derivatives must display "Built with Muse Spark Open" in an allowed location such as the interface, documentation, website, or settings.[3]

Section 3 requires a separate written agreement with Meta before certain Model-as-a-Service offerings. It applies when the licensee or affiliates provide third-party access to AI models through an API or similar interface and their aggregate gross revenue exceeds USD 20 million during any consecutive 12-month period. The clause excludes internal use, research, commercial activities that are not Model as a Service, and use as a component of a product that is not such a service. It is not a blanket USD 20 million limit on every commercial use.[3]

The license incorporates a usage policy and says Meta claims no ownership of generated outputs.[3]

## Safety and limitations

Meta applied additional safety post-training for the open-weight release.[1] Its model card warns about inaccurate or biased responses, reasoning errors, language differences, and quantization effects.[2]

The prompting guide recommends video clips longer than a few seconds at 448 pixels or more on the short side; very short or low-resolution clips can trigger a response that the model cannot view them.[8]

Meta's StrongREJECT v2 test adapts the original benchmark with 202 disallowed behaviors and multiple attack conditions.[9] The original [StrongREJECT](https://aiwiki.ai/wiki/strongreject) research measures whether responses provide useful harmful information, rather than only whether a refusal was bypassed.[12]

The card says results do not automatically transfer to modified models and recommends human confirmation for irreversible actions.[2]

## References

1. Meta Superintelligence Labs. ["Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date."](https://research.meta.ai/blog/muse-spark-open) October 15, 2026.
2. Meta Superintelligence Labs. ["Muse Spark Open Model Card."](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) Released October 15, 2026; accessed October 16, 2026.
3. Meta Platforms. ["Muse Model License,"](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B/blob/main/LICENSE.md) version Muse Spark Open, including Exhibit A usage policy. October 15, 2026.
4. Albert Lee, Pavel Belevich, and Marylise Tauzia. ["Muse Spark Open, an open-weight model for coding agents."](https://www.baseten.co/blog/muse-spark-open-launch/) Baseten, October 15, 2026.
5. vLLM Team. ["Day 0 support for Muse Spark Open on vLLM."](https://vllm.ai/blog/2026-10-15-muse-spark-open) October 15, 2026.
6. Meta. ["Get the model."](https://dev.meta.ai/docs/muse-spark-open/get-the-model) Muse Spark Open developer documentation. Accessed October 16, 2026.
7. Adam Loving et al. ["Build with Muse Spark Open."](https://dev.meta.ai/resources/blog/build-with-muse-spark-open) Meta for Developers, October 15, 2026.
8. Meta. ["Prompting guide."](https://dev.meta.ai/docs/muse-spark-open/prompting) Muse Spark Open developer documentation. Accessed October 16, 2026.
9. Meta Superintelligence Labs. ["Muse Spark Open Evaluation Methodology."](https://research.meta.ai/static/muse-spark-open-evaluation-methodology) October 2026.
10. Tianbao Xie et al. ["OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments."](https://arxiv.org/abs/2404.07972) NeurIPS 2024; arXiv:2404.07972.
11. PyTorch Team at Meta. ["Muse Spark 1.2 Is Now Open Weights."](https://pytorch.org/blog/muse-spark-open) October 15, 2026.
12. Alexandra Souly et al. ["A StrongREJECT for Empty Jailbreaks."](https://arxiv.org/abs/2402.10260) NeurIPS 2024; arXiv:2402.10260.
13. Meta. ["Muse Spark Open on Mac."](https://dev.meta.ai/docs/cookbook/muse-spark-open-mac) Developer cookbook, accessed October 16, 2026.