Meta released Muse Spark Open on October 15, 2026. The open-weight checkpoint is Muse-Spark-Open-816B-A42B: a mixture-of-experts model with 816 billion total parameters and about 42 billion active per token. It takes text, images, video, and audio, and replies in text, with a 1 million token context window at inference.
Weights ship under the Muse Model License. Meta’s research announcement positions the release for coding and agentic work: long-horizon plans, tool use, and multi-step software tasks.
Muse Spark Open is a MoE Transformer with a Perception Encoder for multimodal input. Only a subset of experts runs for each token, so the active compute stays closer to a mid-size dense model while the full expert set still has to fit in memory.
Meta’s model card highlights multi-step reasoning, end-to-end agent tasks, coding agents, tool calling, multimodal reasoning over screenshots and recordings, and related evaluation or data-generation work. Recommended sampling for most deployments is temperature 1.0 and top_p 1.0, with reasoning effort set to high or xhigh for harder coding and agent jobs.
On Meta’s published comparison table (xhigh reasoning), Muse Spark Open leads several agent and knowledge suites among the open-weight peers Meta lists, including MCP-Atlas, HLE-Diamond without tools, JobBench, and AA Omniscience. Other coding suites such as Terminal-Bench 2.1 stay competitive rather than first place.
You can call hosted APIs from partners Meta names, or download the weights yourself. Self-hosting starts from two Hugging Face checkpoints:
Those weight sizes do not include the KV cache, activations, or runtime overhead, so a long 1M-token context still adds working memory on top.
Supported inference frameworks Meta lists include SGLang, vLLM, Hugging Face Transformers, TensorRT-LLM, Unsloth, llama.cpp, and Apple’s MLX.
Meta lists Mac Studio as an official platform for Muse Spark Open. You run the model with Apple’s MLX software on Mac Studio machines with M5 Ultra chips.
One Mac Studio is not enough for Meta’s official Mac pack. Meta’s hardware guide calls for two Mac Studios with 512 GB of unified memory each, or four with 256 GB each. The Macs share the model over Thunderbolt 5. Apple’s machine-learning team published the same setup and measured Muse Spark Open on two- and four-machine clusters.
Meta’s official downloads need far more graphics memory than a single consumer GPU. The model card lists about 700 GB for one official pack and about 2 TB for the other, spread across several high-end GPUs in one machine or node.
Some community builds shrink the files further for desktop apps. Those are a different download, with their own size notes. Size Meta’s official packs from Meta’s figures above.
CORSAIR PRO builds configure-to-order AI workstations, from deskside systems to multi-GPU towers. For Muse Spark Open you can configure a system to meet the requirements to run and load the model.
Explore options in the CORSAIR PRO configurator, or talk through requirements with a CORSAIR PRO solutions architect.
Stay up to date with CORSAIR. Get our latest News, Guides, and Product Updates in your Google feeds.
Add CORSAIR as a preferred sourceThanks for helping us improve this article.
JOIN OUR OFFICIAL CORSAIR COMMUNITIES
Join our official CORSAIR Communities! Whether you're new or old to PC Building, have questions about our products, or want to chat about the latest PC, tech, and gaming trends, our community is the place for you.