Muse Spark Open: 816B Weights, and Who Can Actually Run Them

Meta shipped 816B parameters you can download — and almost nobody can run

On October 15, 2026, Meta Superintelligence Labs published the weights for Muse Spark Open: a sparse mixture-of-experts model with 816B total parameters and roughly 42B active per token, a 1M-token context window, and native input across text, images, audio, and video. It outputs text. The license is the Muse Model License, Meta's own instrument — not Apache 2.0.

The same week, Reflection AI released Beam, a 501B model under Apache 2.0. Two American frontier-scale open-weight releases in seven days is a pattern worth understanding, and the interesting parts are not the parameter counts.

Here is the headline: Muse Spark Open is not a model you run at home. It is a model you rent, or one you run if you already own a GPU estate — or two Mac Studios. The benchmarks are the least important thing in this release. The memory bill and the license terms are what will reach your procurement plan.

What is actually inside

62 layers, hidden dimension 8192. Each MoE layer holds 256 experts and routes each token to the top 8. Experts hold about 97% of all parameters.
GQA with 128 query heads and 16 KV heads, head dimension 64. Untied input and output embeddings, about 1.66B each, with a 202,048-token vocabulary.
Perception encoder: 1.8B parameters, 50 layers, 14×14 patches, 4096 output dims. Every fourth layer plus the last uses full attention; the rest are sparse.
Audio enters at one token per 80 milliseconds — 12.5 tokens per second, so ten seconds of audio costs 125 tokens.
Context is 1,048,576 tokens, and images, video frames, and audio share that window with text.
You cannot turn reasoning off. reasoning_effort accepts low/medium/high/xhigh (internally 16/64/256/512), and the default is xhigh. For cheap tasks, drop it to low rather than expect a non-thinking mode.

The benchmark table, read honestly

Meta's model card reports these at xhigh effort, alongside seven competing open-weight models. Every number is vendor-measured.

Artificial Analysis Intelligence Index: Muse Spark Open 39 — GLM 5.3 45, Kimi K3 44, Qwen3.8-2.4T-A95B 40, Mistral Large 4 38.
Terminal-Bench 2.1: 76.4 — Beam 80.1, Qwen3.8 82.0, GLM 5.3 83.9, Kimi K3 85.0.
DeepSWE 1.1: 60 — Mistral Large 4 62, GLM 5.3 67, Kimi K3 68, Beam 44.
MCP-Atlas: 89.2 — first among the eight. Kimi K3 82.3, GLM 5.3 84.2, Beam 78.7.
JobBench: 61.8 — narrowly ahead of GLM 5.3 at 61.4.
OSWorld Verified: 80.4 — Kimi K3 84.8.
HLE-Diamond, no tools: 26.1, against 22.2 for the next best.
ZeroBench with tools: 50.0 against 41.0 for Kimi K3.

The shape of the result matters more than any single row. On general intelligence, Muse Spark Open loses to GLM 5.3 and Kimi K3. On raw coding it trails too. What it wins is the plumbing of agent work: tool selection and screen control. MCP-Atlas at 89.2 and JobBench at 61.8 say the model is unusually reliable at calling the right tool and finishing multi-step jobs, which is a different skill from writing clever code.

Meta calls it "the most capable US open-weight model to date." Read that sentence twice: the qualifier is doing all the work. Strip "US" and the ranking changes.

What serving it actually costs

Sparse MoE saves compute, never memory. Eight of 256 experts fire per token, so each token costs about 42B parameters of arithmetic — but all 256 experts must be resident.

BF16: roughly 1.5 TB of safetensors. The reference Transformers config runs across four nodes of 8× H200 with tensor parallelism of 32.
NVFP4: one B200 node; Meta lists a 700GB VRAM floor for the 4-bit checkpoint.
KV cache: each token adds about 250 KB (2 × 16 heads × 64 dims × 62 layers at BF16), which is ~266 GB at the full 1M context before weights.
SGLang reference: NVFP4 across 4× GB300.
Hugging Face Inference Endpoints: 8× RTX Pro 6000 (98GB) for a 131K context, or 8× H200 (141GB) for the full million.

Hugging Face's own measurements are the useful part. On an M3 Ultra with 512GB, an imatrix-calibrated 2-bit quant runs at 18 tokens/s with 304GB of model memory. The oMLX 3-bit quant takes about 344GB and runs at roughly 225 tokens/s prefill and 21 tokens/s generation. Two M5 Ultra Mac Studios with 512GB each, linked over Thunderbolt 5 with RDMA, hold the NVFP4 checkpoint with tensor parallelism splitting every layer across both machines.

So "runs on your desktop" means $14,000–20,000 of Mac hardware for 20 tokens/s, or a Blackwell node. For everyone else, this is an API model.

The API price is the one number most teams will use

Across OpenRouter, Baseten, and Vercel AI Gateway the rate is identical: $1.25 per million input tokens, $4.25 per million output, $0.15 cached input. The model ID is meta/muse-spark-open, served by Baseten, Fireworks, Together, Modal, Nebius, Groq, and OpenRouter. OpenRouter's endpoint data showed Together at 91.8% uptime over 24 hours at the time of writing, with Baseten advertising the fp4 variant.

Compare that with Beam at $1.30 input / $0.13 output. Beam's output is roughly a third of Spark's price. If your workload is long generations, Beam is cheaper by a wide margin. If your workload is many tool calls per short answer, Spark's MCP-Atlas lead is the reason to pay.

The license is the actual decision

The Muse Model License grants free use, copying, modification, fine-tuning, distillation, merging, publishing, distribution, and sale — subject to three conditions that a procurement review will ask about.

Attribution. Any product or service using the model or a derivative must prominently display "Built with Muse Spark Open" in the interface, documentation, website, or settings menu. A derivative is defined broadly: anything "trained, modified (including fine-tuned), distilled, or otherwise improved" using the model or its outputs. Fine-tune a smaller model on Spark traces and the notice requirement follows you. The $20 million clause. If you provide third parties access to AI models via an API or similar interface — model-as-a-service — and your group's aggregate gross revenue exceeds $20 million in any consecutive twelve months, you must notify Meta and sign a separate written agreement before offering it. Internal use, research, and powering features of your own product are exempt. This is a conditional grant, not an irrevocable one. Usage policy. No critical infrastructure operation, transportation technologies, heavy machinery, or use by anyone under 18.

None of this makes the release useless. It makes it a grant with a renegotiation priced into it, which is exactly the kind of term an open-source license never contains.

What to do this quarter

Test the number, not the press release. Run your own tool suite through the API and see whether MCP-Atlas 89.2 reproduces. At $1.25/$4.25 an evaluation harness costs tens of dollars.

Start with NVFP4, not BF16. One B200 node is a purchase order. Four 8×H200 nodes is a capital program. Starting with the second is how a pilot quietly dies.

Price the Mac path honestly. Two M5 Ultra Studios deliver about 20 tokens/s. That is a batch and overnight job machine, not an interactive one.

Read the license before the benchmark table. Attribution and the $20M trigger touch product UI and finance, and neither appears in the parameter-count coverage.

Benchmark against Beam. Same week, same class, wildly different license. Beam for freedom and cheap output, Spark for tool reliability.

Two frontier open-weight models landed in one week, and "open" meant something different each time. That is the real story of October: the weights are becoming downloadable faster than the licenses are becoming permissive, and the selection decision has moved from parameter counts to memory budgets and contract terms.