Research note
Research note: Meta releases Muse Spark Open, an 816-billion-parameter open-weight model that can run on two Mac Studios
How we checked Muse Spark Open’s size, licence terms and benchmark claims against Meta’s post and model card.
What we set out to check
What Meta released, how big it is, what the licence allows, and how Meta’s benchmark numbers compare with rival open models.
What we found
- Meta’s post gives 816B total and 42B active parameters, the 1M context and the list of download and API partners.
- The licence on Hugging Face sets a $20 million revenue threshold for hosted services, not a user-count threshold.
- The benchmark table shows Muse Spark Open leading on several knowledge and multimodal rows and trailing Kimi K3 or GLM 5.3 on most coding and agent rows.
Numbers that differ
One summary described the hardware as two to four Mac Studios. Meta’s post says two; we use Meta’s figure.
Labels we chose and why
- Confirmed for release, size and licence terms.
- Company claim for the Mac Studio claim and all benchmark scores.
What we could not verify
- Independent benchmark results.
- The full safety evaluation report, which Meta says is coming.
What we searched
- Meta Muse Spark Open release
- Muse Spark Open Hugging Face licence
- Muse Spark Open benchmarks Kimi K3
Every source we read
| Source | Used? | Why |
|---|---|---|
| Meta AI blog | USED | Size, context, availability, Mac Studio claim, safety statement. |
| Hugging Face model card | USED | Licence terms and the benchmark table. |
| The Register | USED | September report that open weights were coming. |
| AI Weekly | NOT USED | Summary only. |