Muse Spark Open
by meta-models
Meta's open-weights Muse Spark: 816B total and 42B active parameters, text, images, audio and video in, text out, with a 1M-token context
meta-models/Muse Spark OpenDeploy Muse Spark Open
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
Muse Spark Open is the open-weights release of Meta Superintelligence Labs' Muse Spark, published on Hugging Face on 15 October 2026 under the Muse Model License. It is a mixture-of-experts model with 816B parameters in total and 42B active per token, and it reads text, images, video and audio and answers in text. Meta built it for coding and agentic work: multi-step tool use, long tasks, and reasoning over screenshots, charts, documents and recordings. The context window is 1M tokens.
On Meta's own comparison it leads the open models it lists on MCP-Atlas (89.2), JobBench (61.8), ZeroBench with tools (50.0) and SimpleVQA with tools (73.6), and trails Kimi K3 on OSWorld Verified, Terminal-Bench 2.1 and VideoMMEv2. Meta also gave this version extra safety post-training so it holds up without a system guardrail in front of it, though it still recommends one.
Running it yourself takes serious hardware. Meta's card asks for at least 700 GB of GPU memory for the NVFP4 weights and 2 TB for BF16, and says it is optimized to run on two Mac Studios with M5 Ultra using MLX. The weights are gated: you request access on Hugging Face. We have not measured it.
Architecture
A mixture-of-experts transformer with 62 layers, hidden size 8192, gated attention with QK-norm, 128 query heads and 16 key/value heads, and a 202,048-token vocabulary with untied input and output embeddings. A 1.8B-parameter Perception Encoder (50 layers, 1536 hidden size, 14x14 patches, every fourth layer and the last using full attention) turns images and video frames into tokens for the language model; audio is also an input. 816B parameters in total, 42B active per token, 1M-token context at inference. Shipped in BF16 and NVFP4.
Mixpeek SDK Integration
# Muse Spark Open is too large to run per frame. Index the library with Mixpeek's
# extractors, and give the model a search tool so it can pull the right clips itself.
import requests
requests.post(
"https://api.mixpeek.com/v1/buckets/bkt_your_bucket/objects",
headers={"Authorization": "Bearer API_KEY", "X-Namespace": "ns_your_namespace"},
json={"blobs": [{"property": "video", "type": "video", "url": "s3://bucket/calls/2026-10-14.mp4"}]},
)Capabilities
- Reads text, images, video and audio together and answers in text
- 1M-token context for long documents, long recordings and long agent sessions
- Tool calling across long multi-turn tasks, with retries when a tool call fails
- A reasoning_effort setting of low, medium, high or xhigh (the default)
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| MCP-Atlas | Score | 89.2 | Model card: meta-models/Muse-Spark-Open-816B-A42B (self-reported, xhigh effort; Qwen3.8-2.4T-A95B 84.5) |
| OSWorld Verified | Score | 80.4 | Model card (self-reported; Kimi K3 84.8) |
| SimpleVQA (with tools) | Score | 73.6 | Model card (self-reported; Kimi K3 71.9) |
| VideoMMEv2 | Score | 61.5 | Model card (self-reported; Kimi K3 65.4) |
| GPQA Diamond | Accuracy | 89.0 | Model card (self-reported; Kimi K3 93.5) |
Performance
Supported runtimes per the card: SGLang, vLLM, Hugging Face Transformers, TensorRT-LLM, Unsloth and llama.cpp, plus MLX on two M5 Ultra Mac Studios. Recommended sampling is temperature 1.0, top_p 1.0, top_k disabled. Meta's card says videos of a few seconds at around 100 px can be refused as unviewable; use 448 px or more on the short side. We have not measured it.
Common Pipeline Companions
The hosted Muse Spark on Meta's API, when you would rather not run 816B parameters yourself
Embeds the library so a retriever can find the clips and pages the model should read
A small captioner for every frame, leaving Muse Spark Open for the questions that need it
Frequently Asked Questions
What is Muse Spark Open?
The open-weights version of Meta's Muse Spark, released on 15 October 2026. It is a mixture-of-experts model with 816B total and 42B active parameters that reads text, images, video and audio, writes text, and has a 1M-token context. Meta built it for coding and agentic work.
What hardware does Muse Spark Open need?
Meta's model card asks for at least 700 GB of GPU memory for the NVFP4 weights and 2 TB for BF16. It also says the model is optimized to run on two Mac Studios with M5 Ultra using MLX, and community 3-bit and 4-bit MLX and GGUF quantizations are on Hugging Face.
Is Muse Spark Open open source?
The weights are downloadable under the Muse Model License, a custom license with a usage policy, and access on Hugging Face is gated by request. Read its terms before commercial use, because they differ from Apache 2.0 or MIT.
Can Muse Spark Open understand video and audio?
Yes. Its inputs are text, images, video and audio. Meta reports 61.5 on VideoMMEv2 and warns that very short or very small clips (a few seconds at around 100 px) may be refused, so use 448 px or more on the short side.
How do I use Muse Spark Open with Mixpeek?
Let Mixpeek index and search the library, and give the model a tool that runs a retriever, as in the example on this page. The model then pulls the clips, frames and documents it needs into its context instead of reading the whole archive.
Specification
Research Paper
Build with Muse Spark Open (Meta developer blog)
arxiv.orgBuild a pipeline with Muse Spark Open
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free