# Meta Releases Muse Spark Open Weights: An 816B Mixture-of-Experts Under Attribution and $20M Model-Service Terms

Times of AI Desk · 2026-10-16 · Models

[https://timesof.ai/2026/10/meta-muse-spark-open-816b-open-weights](https://timesof.ai/2026/10/meta-muse-spark-open-816b-open-weights)

> Meta released weights for Muse Spark Open on Oct. 15, an 816B-total, 42B-active multimodal mixture-of-experts with a 1M-token context, which Meta calls 'the most capable US open-weight model to date.' The Muse Model License requires a 'Built with Muse Spark Open' notice and a separate deal for large model-as-a-service providers. Benchmarks are Meta's own.

Meta is back to large open weights, and by its own table it still trails some Chinese open models on coding. **Meta** on **Thursday, October 15** released the weights for **Muse Spark Open**, which it calls "the most capable US open-weight model to date." That is Meta's claim.

## What the model card describes

Muse Spark Open is a mixture-of-experts transformer with **816 billion** total and **42 billion** active parameters, 256 experts with eight active per token, and a **1M-token** context, according to its model card. It takes text, image, video and audio input and outputs text. Meta says the open-weight version received additional safety post-training for use without system-level guardrails.

Weights come in BF16 and NVFP4 on Hugging Face, Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry, and seven hosts offer API access, Meta says. Running it yourself needs at least 700GB of GPU memory for NVFP4 or 2TB for BF16, according to the card, which says it is also optimized to run on two M5 Ultra Mac Studios using MLX.

## Meta's benchmarks

In Meta's own comparison table, Muse Spark Open leads Kimi K3, GLM 5.3 and Qwen3.8 on agentic tests such as MCP-Atlas (89.2) and JobBench (61.8), but trails them on Terminal-Bench 2.1 (76.4 against Kimi K3's 85.0) and DeepSWE. The same table lists an Artificial Analysis Intelligence Index score of 39, against 45 for GLM 5.3, 44 for Kimi K3 and 40 for Qwen3.8. Meta's safety figures, such as refusing 99.2% of jailbreaks on its internal cyber benchmark, are also its own tests; a full evaluation report is "available soon."

## The licence

The **Muse Model License** is permissive but requires licensees to "prominently display 'Built with Muse Spark Open'" in products using the model or its derivatives. Model-as-a-service providers whose group revenue exceeds $20 million in any 12-month period must notify Meta and sign a separate agreement. Meta released its closed [Muse Wonder](https://timesof.ai/2026/10/meta-muse-wonder-first-frontier-model) days earlier.

## Limits

- Benchmark and safety figures are Meta-reported; the AA index score is as listed on Meta's card. Reports that a Meta executive said this is not its most powerful model come from relays of CNBC and WSJ that were not read, so they are not included.

## Sources

- [Meta AI Research: Introducing Muse Spark Open (October 15, 2026)](https://research.meta.ai/blog/muse-spark-open)
- [Hugging Face: Muse Spark Open model card and Muse Model License](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B)
- [Times of AI: Meta releases Muse Wonder (October 13, 2026)](https://timesof.ai/2026/10/meta-muse-wonder-first-frontier-model)