
None
None
Long Story Video Skill
Turn your idea into 10s–10min video—AI writes scripts, prompts, and generates footage automatically.

Ads Video Skill
Generate professional ads and sales videos—AI auto-generates scripts, prompts, and footage.
3D Science Explainer Video Skill
Convert scientific concepts into stunning 3D explain animations
Feedback
freeTrialImage.bannerPity
freeTrialImage.upgradeUnlock
- ✓freeTrialImage.benefitHd
- ✓freeTrialImage.benefitWatermark
- ✓freeTrialImage.benefitUnlimited
Muse Spark Open
Muse Spark Open explained: 816B open-weight parameters, a million-token window, true multimodality and local video agents on Apple silicon.
All Tools
Discover our comprehensive AI-powered animation toolkit
MiniMax H3
MiniMax H3 AI Video Generator
Seedance 2.5
The Future of AI Video Is Here.

Seedance 2.0
The Future of AI Video Is Here.

Seedance 2.0
The Future of AI Video Is Here.

Gemini Omni
Gemini Omni Video Generator
Image to Video
Generate videos from images

AI Multi-Scene Shorts Generator
Create viral AI Shorts instantly

3D Science Video
Create 3D science videos easily
Inside Muse Spark Open: 816B Parameters, Zero Cloud Lock-In
Shipped as downloadable weights, Muse Spark Open is a multimodal engine built for coding, autonomous agents and lengthy multi-step jobs — and it answers to hardware you already own.
- Downloadable Weights, Built in the USWith 816 billion total parameters and 42 billion active per token, this release pairs a million-token window with native multimodality — the most capable open-weight drop to date from a US lab.
- Workstation-Scale IntelligenceCapability this large usually hides behind a data-center wall. Here it boots on local hardware: two Apple M5 Ultra Mac Studios running the newest MLX build carry the whole thing.
- Agents That Already Ship Real WorkRunning through Muse Code, the model pushed genuine iOS patches into Ice Cubes and then operated DaVinci Resolve unaided, picking the strongest frames and cropping them vertically.
Three Ways to Get Muse Spark Open Running
Choose your path: spin it up on Apple silicon, serve it with the engine you prefer, or call a managed endpoint and skip the setup entirely.
What Muse Spark Open Can Do
Million-token memory, true multimodality, workstation bandwidth and weights you control: the ingredients that let autonomous video editing actually reach the finish line.
A One-Million-Token Window
Long agent runs never lose the thread — a million tokens of history stay in view, so multi-step reasoning keeps earlier decisions on hand.
Built for Long Agent Runs
Trained on real engineering loops, it stays on task across hours of work and can drive applications like DaVinci Resolve from first click to final render.
Multimodal by Design
Images, audio and text enter the same network rather than passing through bolt-on adapters, which is exactly what lets it inspect footage frame by frame.
Two Macs, One Pooled Machine
Each M5 Ultra GPU reaches up to 512 GB of unified memory, so a pair of machines address roughly a terabyte together and hold the entire model as a single unit.
1.2 TB/s of Memory Bandwidth
The M5 Ultra shuttles 1.2 TB/s, plenty to feed the small active-parameter slice behind every token and keep responses snappy on a local rig.
Hardened Before Release
Open releases get extra safety tuning. The model was benchmarked against the Meta Superintelligence Scaling Framework and passed its launch threshold in every risk category measured.
Muse Spark Open: Your Questions Answered
Straight answers on hardware, licensing, downloads and API access for this open-weight release.
Which setup can actually host the model locally?
Two Apple M5 Ultra Mac Studios on the latest MLX build. Each addresses up to 512 GB of unified memory, so the pair offers around a terabyte — enough for the complete checkpoint.
How big is the model, and how much context fits?
Expect 816 billion parameters overall with 42 billion active per token, a million-token context, and multimodality baked in — tuned for software work, autonomous agents and extended tasks.
Where do I get the weights, or call it through an API?
Grab the files from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden or Microsoft Foundry Models. Prefer managed endpoints? Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI all serve it.
Which engines and chips does it support?
Pick from MLX, llama.cpp, Unsloth, SGLang, vLLM, TensorRT-LLM or Transformers. It is tuned across NVIDIA, AMD, Intel and Apple silicon, with PyTorch available underneath for deeper customization.
What license covers the released weights?
The Muse Model License governs use. Its limited commercial terms still allow the large majority of developers to run the model at no charge.
How was safety checked before launch?
Before shipping, it was measured against the Meta Superintelligence Scaling Framework for the most severe risk areas, received additional post-training for open distribution, and passed the launch bar — the model card has the full breakdown.
Start Building With Muse Spark Open
Pull the checkpoint from Hugging Face, point your app at a managed endpoint, then show the community what you made in the forums.
