Muse Spark
First Meta Superintelligence Labs model, a closed, natively multimodal reasoning model on a rebuilt stack and Meta's first flagship without open weights.
New architecture, data and infra rebuilt from scratch; Meta says it reaches Llama 4 Maverick capability with over an order of magnitude less pretraining compute. Adds visual chain of thought, tool use, and a Contemplating mode that runs parallel agents. Larger models said to be in development.
- Date
- Wednesday, 8 April 2026
- Lab
- Meta
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Humanity's Last Exam (Contemplating mode) | 58% FrontierScience Research 38% | company |
| Pretraining compute vs Llama 4 Maverick | >10x less for equivalent capability, per Meta | company |
Product in Meta AI app and meta.ai; private API preview for select partners at launch. Codename Avocado (reported). Third-party index comparisons were not opened and are omitted.
Sources
- ai.meta.com/blog/introducing-muse-spark-msl/
- www.latent.space/p/ainews-meta-superintelligence-labs
- en.wikipedia.org/wiki/Muse_Spark
This record was partly confirmed: some claims could not be checked on 6 October 2026. How we check