Muse Spark Open: Wait for Benchmarks Before Committing
Meta's 816B open-weight model ships with specs and hardware numbers but no benchmarks or license on its launch page.
The bottom line: Wait. Muse Spark Open is a real open-weight release, but Meta's Hugging Face post gives no benchmark scores and no license, so its "most capable US open-weights model" claim is unverified. At about 816B parameters it needs multi-node GPUs or 256GB-plus RAM. Only well-resourced labs should test it now.
The Weights Desk · 4 min read- Verdict: Wait. The launch page gives architecture and hardware figures but no benchmark scores and no stated license.
- The model has about 816B total and about 42B active parameters, with 256 experts per layer and 8 selected per token.
- Full BF16 deployment needs four 8-GPU H200 nodes. The 3-bit MLX build needs more than 256 GB of RAM.
- Reasoning cannot be switched off, so simple tasks still spend reasoning tokens.
- Meta's earlier flagship, Muse Spark, launched closed, so the license of this release is the first thing to check.
Our verdict on Muse Spark Open is Wait. Meta's open-weight release, announced in a Hugging Face post dated October 15, 2026, publishes detailed architecture and deployment figures. It does not publish benchmark scores or state a license, and those are the two facts that decide whether a team should build on it. The headline number is about 816 billion total parameters, with about 42 billion active per token.
What the release actually contains
The post describes a mixture-of-experts model with 62 layers, 256 experts per layer and 8 selected per token, with experts holding about 97% of parameters. It accepts text, images, audio and video, and outputs text only. Context runs to 1,048,576 tokens. Checkpoints come in BF16 (about 1.5 TB) and NVFP4, with community GGUF and MLX quantizations also listed. These are Hugging Face's own figures, not independently measured.
The benchmark and license gap
The post calls the model the most capable US open-weights model to date, yet the text we could read offers no benchmark scores, no baseline and no harness to reproduce a comparison. It also states no license. For context, Artificial Analysis described the original Muse Spark as Meta's first non-open release with a proprietary license, so the terms attached to these weights are a real open question, not a formality.
Hardware is the real filter
Running it at full precision takes four nodes of eight H200 GPUs in Transformers. The NVFP4 build fits a single B200 node, and vLLM requires Blackwell GPUs. On a 512 GB M3 Ultra, the post reports about 344 GB used and roughly 21 tokens per second for a 3-bit MLX build, and says that build needs more than 256 GB of RAM. A full 1M-token KV cache alone is about 266 GB in BF16.
Where it breaks
The post itself lists the limits. Reasoning cannot be turned off; effort levels run from low to xhigh, with xhigh the default, so even trivial requests spend reasoning tokens. Quantized builds need more prompt steering. The fine-tuning example freezes the experts and trains LoRA adapters on attention projections, using 64 H100 GPUs for about two hours. Token accuracy rose from 72% to 84% on that example, which says little about general quality.
The bottom line
The bottom line: Wait. The number that decides it is zero, the count of benchmark scores on the launch page. Labs and large enterprises with multi-node GPU capacity, a legal team to read the license once confirmed, and their own evaluation set can test it now against a stated baseline. Everyone else should hold until independent results and license terms exist, because a launch-day superlative with no numbers behind it is a lead, not evidence.
- Is Muse Spark Open actually open-weight?
- Hugging Face's launch post lists downloadable BF16 and NVFP4 checkpoints under the meta-models organization. The post does not state a license, so reuse and commercial terms must be read from the model repository before anyone builds on it.
- Can a small team run Muse Spark Open?
- Only in heavily quantized form. The post reports the MLX 3-bit build using about 344 GB on a 512 GB M3 Ultra at roughly 21 tokens per second. The BF16 build needs four nodes of eight H200 GPUs.
- Is it the best US open model?
- Unverified. The post calls it the most capable US open-weights model to date but gives no benchmark scores or baseline comparison in the text we could read.
- Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta — Hugging Face
- Muse Spark: everything you need to know — Artificial Analysis