# Meta releases Muse Spark Open, a 816 B multimodal MoE model

> The new open-weight model can handle text, images, audio and video, runs with 42 B active parameters per token and supports up to a million‑token context.

Oossa · 2026-10-15 · https://oossa.com/en/meta-releases-muse-spark-open-a-816-b-multimodal-moe-model

Meta announced Muse Spark Open on 15 October 2026. It is an open‑weight mixture‑of‑experts (MoE) model that can process text, images, audio and video in a single prompt. The model has about 816 billion total parameters, but only roughly 42 billion are active for each token, letting it run faster than a full‑size model of that scale. It also offers a massive context window of up to one million tokens, meaning long documents, codebases or video transcripts can be handled in one go.

## What the model can do

Muse Spark Open is tuned for “agentic coding” – reading a codebase, writing or debugging code, and calling external tools while looking at screenshots, voice notes or screen recordings. It generates only text, but the input can be any combination of modalities. The model ships in two checkpoint formats: BF16 (about 1.5 TB) and a 4‑bit NVFP4 version that runs on NVIDIA Blackwell GPUs. Meta provides integration recipes for llama.cpp, vLLM, SGLang and other inference frameworks, plus a hosted Hugging Face Inference Endpoint that uses an OpenAI‑compatible API.

## The facts

- Muse Spark Open has ~816 billion total parameters and ~42 billion active per token.
- The model uses 256 experts per MoE layer, with 8 experts selected for each token.
- It accepts text, images, audio and video as inputs and outputs text only.
- Context window can hold up to 1 000 000 tokens, which requires about 266 GB of KV cache in BF16.
- The BF16 checkpoint is ~1.5 TB; the NVFP4 4‑bit checkpoint runs on a single NVIDIA H200 node.

## Why it matters

Developers can now run a single AI model that understands code, screenshots and voice notes without switching tools, potentially speeding up debugging and documentation tasks. However, the model still needs high‑end GPU hardware or a hosted endpoint, so it remains out of reach for casual users without such resources.

## Sources & references

1. [Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta](https://huggingface.co/blog/meta-muse-spark-open) – Hugging Face, 2026-10-15
2. [New on OpenRouter: Meta: Muse Spark Open](https://openrouter.ai/meta/muse-spark-open) – OpenRouter API, 2026-10-15
3. [Muse Spark 1.2 Is Now Open Weights](https://pytorch.org/blog/muse-spark-open/) – PyTorch
4. [Meta launches an open-weight AI model that runs on Macs](https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/) – TechCrunch, 2026-10-15
5. [Muse Spark Open is Meta's answer to Qwen, GLM, and Kimi in the open-weight race](https://the-decoder.com/muse-spark-open-is-metas-answer-to-qwen-glm-and-kimi-in-the-open-weight-race/) – The Decoder, 2026-10-15
6. [Muse Spark Open, an open-weight model for coding agents](https://www.baseten.co/blog/muse-spark-open-launch/) – Baseten
7. [Apple and Meta team up to make Muse AI model run on Mac computers](https://www.cnbc.com/2026/10/15/apple-meta-partner-ai.html) – CNBC, 2026-10-15
8. [Meta and Apple team up to make Muse AI models run better on Macs](https://www.theverge.com/news/1012293/meta-apple-muse-spark-open-ai-model-mlx) – The Verge, 2026-10-15

Last updated: 2026-10-15