# Meta's Muse Spark Open brings a frontier MoE to open weights

> The 816B-parameter mixture-of-experts model activates just 42B per token and targets agentic coding across a million-token context.

Published by The Open Weights on Sep 15, 2026. Canonical: https://theopenweights.com/news/muse-spark-open-816b-a42b-41dr

## Key facts

- Company: Meta AI
- Model: Muse Spark Open 816B-A42B
- Version: Muse Spark Open
- Category: Any-to-Any
- Modalities: Text / LLM, Any-to-Any, Vision-Language
- License: Other (Open weights, commercial use allowed)
- Parameters: 816B (mixture of experts)
- Active parameters: 42B
- Context window: 1M tokens
- Architecture: MuseSparkForConditionalGeneration
- Precision: BF16
- Weights size: 1644.3 GB
- Significance: major
- Published: 2026-09-15
- Last verified: 2026-10-16
- Hugging Face: https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B
- Canonical URL: https://theopenweights.com/news/muse-spark-open-816b-a42b-41dr

Meta AI has released [Muse Spark Open](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B), an open-weight multimodal model built on a mixture-of-experts design. The model totals 816 billion parameters but activates only 42 billion for any given token, a structure meant to deliver frontier-scale capability without the full compute cost of a dense model of the same size.

The headline capability is reach. Muse Spark Open supports a context window of up to one million tokens and handles text and vision inputs, positioning it for agentic coding workflows where a model needs to keep an entire codebase, long logs, or multi-step task histories in view at once.

## Why it matters

Open releases at this scale remain rare. Most models that compete at the frontier stay behind APIs, so weights that developers can inspect, fine-tune, and self-host change what independent teams can build.

- 816B total parameters, 42B active per token
- Multimodal: text and vision inputs
- Up to 1,000,000-token context
- Oriented toward agentic coding

The sparse-activation approach is the practical hook here: it lets a very large parameter count run with the inference footprint closer to a mid-size dense model. Licensing is listed as a custom "other" license rather than a standard permissive one, so teams planning production use should read Meta's terms closely. Full details are available on the [model's Hugging Face page](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B).

## Get the model

- [Hugging Face](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B)

## Sources

- [meta-models/Muse-Spark-Open-816B-A42B](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) — Hugging Face, Sep 15, 2026

---
Source: The Open Weights (https://theopenweights.com/). Aggregated and written by Claude, curated by humans. Cite as: "Meta's Muse Spark Open brings a frontier MoE to open weights", The Open Weights, Sep 15, 2026, https://theopenweights.com/news/muse-spark-open-816b-a42b-41dr