Muse Spark Open: The Best US Open-Weight Model Is a Claim With Terms and Conditions

The claim arrived on X on the evening of October 15, 2026, and it is the kind of claim that is true in one sentence and negotiable in the next. Meta published the weights for Muse Spark Open — the 816-billion-parameter, 42-billion-active checkpoint it had been sitting on since the repository appeared on Hugging Face a month earlier — and within hours the phrase attached to it was "best US open-weight model." The number that most directly supports it is not Meta's. It is 39: the score Artificial Analysis' own hosted listing gives Muse Spark Open at its xhigh reasoning setting, ninth among the 118 open-weight models larger than 150 billion parameters on that board, against a class median of 19. That is a real third-party measurement of a real downloadable model. Separately, Muse Spark 1.3 — the current checkpoint of the closed Muse Spark line, generally available since September 2, 2026 — scores 48 on the same index. Both numbers belong to Meta. Only one of them belongs to something you can download, and only one of them is measured by somebody other than Meta.

The dates, because the framing turns on them

Nothing here is a launch, and the confusion is easy to explain. Meta Superintelligence Labs introduced the Muse Spark family on April 8, 2026. Muse Spark 1.1 arrived July 16, Muse Spark 1.2 on August 5, and Muse Spark 1.3 on September 2 — the last of which picked up its "max" reasoning level on September 4 after additional safety testing. That line is proprietary and served. Muse Spark Open is a separate artifact: a gated Hugging Face repository, meta-models/Muse-Spark-Open-816B-A42B, whose creation timestamp is 2026-09-15, plus an NVFP4 build created seconds earlier. The weights sat there for four weeks inside an access-controlled form. On October 15, 2026, Meta published the licence text, a research blog post, a developer build guide and a benchmark scorecard, and the repository stopped being an artifact nobody had announced.

So the correct reading is neither "stealth drop" nor "launch." The model existed quietly for a month; the paperwork caught up on October 15. What changed for anyone deciding what to run is narrow and specific: there is now a license you can read, a stated hardware recipe, day-one support in the frameworks people actually use, and one independent measurement. What did not change is the closed line, the price of the closed line, or anything about the API models that most readers of this site call.

"Open" is the part with the conditions

The licence is where the headline claim gets its fine print, and it is worth reading before repeating it. Muse Spark Open ships under the Muse Model License. Meta describes it as allowing many developers to use the model for free, and two clauses define where "free" stops.

• Model-as-a-service — if you or your affiliates give third parties access to the model as a product or service over an API, and aggregate gross revenue exceeds twenty million US dollars in any consecutive twelve-month period, you must notify Meta and sign a separate written agreement before running that business. Building the model into your own product is carved out; reselling access to it is not.

• Attribution — the licence requires "Built with Muse Spark Open" to be displayed prominently in the interface of any product that uses the model or a derivative, or in the accompanying documentation, site or settings menu.

• What you may plainly do — fine-tune, distill, merge, publish, sublicense and sell derivatives, with Meta claiming no ownership of outputs.

• How you get it — the repository is gated. The request form asks for full legal name, date of birth and organisation with corporate identifiers, and cannot be edited after submission.

That combination is what makes the "best US open-weight model" phrasing a category claim rather than an absolute one. On the same Artificial Analysis panel, the strongest permissively licensed US open-weight entry is NVIDIA Nemotron 3.5 Super VL at 26.49, with Inkling Small at 25.66 behind it. Muse Spark Open scores higher than both and is not in the same adoption category: a bespoke commercial licence behind a request form is a different thing from a checkpoint you can clone and deploy under OpenMDW-1.1 this afternoon. Meta's own blog post uses the phrasing "most capable US open-weight model to date," which is defensible on the index and narrower than "best" implies once the licence is in the sentence.

What the checkpoint settles without Meta's help

This part is verifiable from the repository itself, before any benchmark is believed.

• Scale — roughly 816 billion total parameters with about 42 billion active per token, a sparse mixture of experts over 62 layers and a hidden dimension of 8,192, with 256 experts per MoE layer and 8 selected per token. Grouped-query attention with 128 query heads and 16 key-value heads at 64 dimensions each.

• Context — 1,048,576 tokens, the window Meta writes as 1M, shared across text, image, video and audio tokens in one budget.

• Modalities — text, images, video and audio in; text out. It reads multimodal input and writes text, which is not the same as generating any of them.

• Checkpoints — a BF16 build of roughly 1.5TB of safetensors, and an NVFP4 4-bit build Meta says runs on a single B200 node. Third parties have since added their own quantisations, from GGUF builds to MLX 3-bit and 4-bit conversions dated the day of release.

• Reasoning control — a reasoning-effort setting with low, medium, high and xhigh levels. Meta's tooling notes that turning reasoning off is not an option.

• Runtimes — Transformers, llama.cpp, SGLang, vLLM, TensorRT-LLM, Apple MLX and Unsloth on day one, with custom auto-tunable Helion kernels for SGLang that Meta reports as kernel-level speedups of 1.03x to 3.95x across NVIDIA and AMD silicon.

• The desktop demo — two Mac Studio machines with M5 Ultra chips under MLX, which works because each exposes up to 512GB of unified memory at 1.2 TB/s. Meta's demos were driving real changes into an open-source iOS client and running DaVinci Resolve end to end.

None of that is a marketing claim; it is a configuration file, a file listing and a published recipe. It is also where the practical cost of "open" lands — the memory arithmetic does not care which licence you accepted. At a full 1M-token context in BF16 the KV cache alone runs to roughly 266GB, so the window the model advertises and the window you can afford on your hardware are two different budgets.

The one number Meta did not write

Artificial Analysis' listing is the only third-party measurement of Muse Spark Open in the public record right now, and it says more than the index alone.

• Intelligence — an Artificial Analysis Intelligence Index of 39 at the xhigh setting, rank 9 of 118 open-weight models above 150B parameters, above a class median of 19.

• Speed — 103.4 output tokens per second, rank 18 of 76 in that cohort.

• Cost — $0.84 per completed Intelligence Index task, on a hosted route priced at $1.25 per million input tokens and $4.25 per million output tokens.

• Efficiency — the listing marks the model as fairly concise, generating roughly 110 million tokens across the index run against a median of 130 million.

Read those against Meta's own scorecard and the gap is not in the specification, it is in coverage. Meta's card reports Terminal-Bench 2.1 at 76.4 with the Terminus 2 agent, MCP-Atlas at 89.2, GDPval-AA v2 at 1401, SciCode at 58.5, JobBench at 61.8, HLE-Diamond at 26.1 with no tools, and DeepSWE 1.1 at 60 on what the card calls the best available harness. Every one of those rows was produced inside the model's own agent harness rather than a shared one, which makes them a capability claim rather than a like-for-like ranking, and we have reproduced none of them. The agreement between the two accounts is on the spec sheet — 816B total, 42B active, 1M context, the same licence, the same input modalities. The disagreement is that one side has eight benchmarks and the other has one index, one speed figure and one price. There is no audio measurement and no video measurement anywhere in public yet, despite the modality row being the most distinctive thing about the model.

Then there is the evaluation Meta has not published.

The launch post says Muse Spark Open was assessed under the Meta Superintelligence Scaling Framework and "met the thresholds our framework requires for release," with detailed results in the model card; the model card points to a full evaluation report described as available soon. A safety claim that cleared an internal framework is a different object from a safety report a reader can inspect, and only the first one exists as of writing.

What this does not change for anyone with a deadline

Muse Spark Open is not something you point an API key at. It is a download, a licence, a gated request form and a hardware decision, and if the workload needs an answer this week the surface area before the first token is large. The practical question is which Muse model you can reach right now, and the answer sits in the closed line: Muse Spark 1.2 and Muse Spark 1.1 are the two checkpoints OrcaRouter routes, both served at Meta's list price of $1.25 per million input tokens and $4.25 per million output tokens with 0% markup, because the provider's list price is passed through rather than marked up. They accept text, image, video, audio and PDF input inside the same 1,048,576-token window, so the multimodal behavior the open checkpoint is being praised for is available on the served line today without a download queue.

Muse Spark Open is not on our catalogue and we are not going to suggest otherwise — a model whose entire proposition is that you can run it yourself is a model you run yourself. Where a routed setup still helps is around it: one API across 200+ models means an unproven checkpoint does not have to be a production bet, automatic failover can carry a request to a second provider when the first degrades, and the routing DSL moves traffic between models by configuration rather than by a code change. For a model released one day ago with a single independent benchmark and an unpublished evaluation report, that is the cheapest way to find out what it does on your traffic while somebody else's infrastructure absorbs the variance.

What would move this piece

Three things are open rather than unknown, and each would change the read. The first is Meta's evaluation report, which is the difference between a safety claim and a safety record. The second is a second independent measurement — one analyst with one index is a data point, not a consensus, and nothing has yet put Muse Spark Open through an audio or video evaluation despite those being its advertised differentiators. The third is how the licence behaves in practice: the twenty-million-dollar threshold and the attribution clause are the terms that will determine whether the strongest US open-weight model by score becomes the most used one, or stays the one teams download to test and then serve from somewhere else. Until at least one of those resolves, the honest version of the headline is that Meta shipped the most capable US open-weight model by the only independent measure available, under a licence that keeps the word "open" specific rather than absolute.

What the numbers cannot tell you is whether any of it has gone wrong in the field: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.