Muse Spark Open: Meta's Flagship Weights Are Out — the Numbers Behind Them Aren't

Muse Spark Open is real, it is downloadable, and it is the largest set of open weights a US lab has published. The repository behind it — M​use-Spark-Open-816B-A42B — first appeared on Hugging Face on September 15, 2026 and sat there, gated and unannounced, for a month. On October 15, 2026 M​eta published the weights, a launch post, a benchmark scorecard and a licence, and the whole thing became official in a single afternoon. So the honest framing is neither "stealth drop" nor "launch": the artifact existed quietly for four weeks, and then the paperwork caught up. What has not caught up is the measurement. A day after release, the vendor's numbers are published and unreproduced, the independent numbers are partial, and the safety evaluation M​eta says backs the release is not out yet. This is what the public record actually supports today.

A month of silence, then everything at once

The dates matter, because they change how you read the release. The Hugging Face repositories — the BF16 checkpoint and a second NVFP4 build — carry creation timestamps of 2026-09-15. The last modification is 2026-10-15, and the licence text inside the gated form is dated October 15, 2026, with the usage policy dated the same day. M​eta's launch post carries a publication timestamp of 2026-10-15. So the sequence is: repo in September, announcement in October.

That is a different story from the one the phrase "shipped quietly" usually describes. M​eta did announce this, in three places at once — its research blog, a developer build guide on its developer site, and a co-authored post on the Hugging Face blog written by Hugging Face staff. What was absent was the press cycle. Roughly a week of launch coverage has not materialised; outside the vendor's own pages and one or two aggregator posts, there is very little written about a model this size. The quiet part is the reception, not the release.

What the repository confirms on its own

Everything below is verifiable from the checkpoint metadata, the configuration, the model card and M​eta's launch post without taking anyone's word for a benchmark. This is the part of the record that is settled.

• Scale — approximately 816 billion total parameters, with roughly 42 billion active per token.

• Architecture — a sparse mixture of experts: 62 layers, hidden dimension 8,192, 256 experts per MoE layer with 8 selected per token, and experts holding about 97% of total parameters. Grouped-query attention with 128 query heads and 16 key-value heads at 64 dimensions each.

• Context — 1,048,576 tokens, the window M​eta abbreviates to 1M, covering text, image, video and audio tokens in the same budget.

• Inputs and output — text, images, video and audio in; text out. This is a multimodal reader, not a generator.

• Checkpoints — two formats. BF16, roughly 1.5TB of safetensors, and NVFP4, a 4-bit build aimed at N​VIDIA Blackwell that M​eta says runs on a single B200 node.

• Memory arithmetic — grouped-query attention keeps the cache small for a model this size, but each token still adds 126,976 values. In BF16 that is about 250KB per token, or roughly 266GB for a full 1M-token context.

• Reasoning — a reasoning-effort setting with low, medium, high and xhigh levels. M​eta's own tooling notes that turning reasoning off is not possible.

• Runtimes — day-one support in Transformers, llama.cpp, SGLang, vLLM, TensorRT-LLM, Apple MLX and Unsloth, with M​eta shipping custom auto-tunable Helion kernels for SGLang that it says deliver 1.03–3.95× kernel-level speedups across N​VIDIA and AMD silicon.

• Desktop — the headline demo. M​eta ran it on 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 of bandwidth, so two machines hold a full BF16 checkpoint. M​eta's demos were shipping real changes to an open-source iOS client and driving DaVinci Resolve end to end.

The licence is the part most people will skip

Muse Spark Open ships under the M​use Model License, which M​eta describes as allowing many developers to use the model for free. Two clauses inside it change what "free" means in practice, and neither is in the headline.

The first is Section 3, the model-as-a-service term. If you or your affiliates provide third parties with access to the model as a product or service over an API, and your aggregate gross revenue exceeds twenty million US dollars in any consecutive twelve-month period, you must notify M​eta and enter a separate written agreement before running that business. Building it into your own product is carved out; reselling the model is not.

The second is attribution. The licence requires licensees to prominently display "Built with Muse Spark Open" in the interface of any product that uses the model or a derivative, or in the accompanying documentation, site or settings menu.

Two things the licence grants plainly, and they are worth stating because they are the actual point of open weights: you may fine-tune, distill, merge, publish, sublicense and sell derivatives, and M​eta claims no ownership of the outputs. M​eta also gates the download itself — the repository is access-controlled, and the request form asks for your full legal name, date of birth and organisation with corporate identifiers, and cannot be edited after submission. Anyone expecting a one-click download should plan for a queue.

What M​eta has claimed but not yet published

M​eta shipped a benchmark scorecard with the release, and the pattern across it is more useful than any single row. The figures below are vendor-reported, each model on the card was measured inside its own agent harness rather than a shared one, and we have not reproduced any of them. They are a claim about capability, not an audited ranking.

• Terminal-Bench 2.1, Terminus 2 agent — 76.4

• MCP-Atlas — 89.2

• GDPval-AA v2 — 1401

• SciCode — 58.5

• JobBench — 61.8

• HLE-Diamond, no tools — 26.1

• DeepSWE 1.1 — 60, self-reported on the best available harness

The strongest rows are agentic and tool-use evaluations; the weakest is closed-book reasoning. That is consistent with what M​eta says the model is for — long-horizon coding, reading a codebase, debugging, calling tools over long runs — and inconsistent with reading it as a general reasoning upgrade over anything.

Then there is the evaluation M​eta has not published. Its launch post says the model was assessed under the M​eta Superintelligence Scaling Framework, that it "met the thresholds our framework requires for release", and that the detailed results sit in the model card. The model card points to a full evaluation report that is described as available soon. As of writing it is not out. A safety claim that has cleared an internal framework is a different object from a safety report a reader can inspect, and only the first one exists today.

The single independent number, and the gap around it

Artificial Analysis has a listing for Muse Spark Open, and it is the only third-party measurement in the public record right now. It puts the model at 39 on the Artificial Analysis Intelligence Index, ninth of 118 entries in its class of open-weight models larger than 150 billion parameters, above the class median of 19. It also measures output speed at 103.4 tokens per second and prices the hosted route at $1.25 per million input tokens and $4.25 per million output, at $0.84 per Intelligence Index task. That listing is worth more than the index: it means the throughput and cost columns that were blank at release are now filled.

Read the two sets of numbers together and the gap is not in the specification, it is in how much of it has been exercised. The analyst's listing independently records 816B total and 42B active parameters, a 1M-token context window, Hugging Face as the weight source and the M​use Model License — the same architecture, the same licence and the same modalities M​eta claims, described as taking text, image, speech and video in and text out. Where the two accounts diverge is coverage. M​eta reports eight benchmarks, all under its own harness. The independent record is one index, one speed figure and one price. There is no audio measurement and no video measurement anywhere in public, no throughput comparison against the closed Muse Spark line, and no second analyst to check the first. Agreement on the spec sheet is not the same as agreement on the numbers, and only the numbers here have been tested.

What running it yourself actually costs

The relevant cost question for open weights is hardware, and here the published details are unusually concrete.

• BF16, datacentre — M​eta's own Transformers recipe runs the BF16 checkpoint across four nodes of eight H200 GPUs with tensor parallelism of 32. That is 32 accelerators for one model.

• NVFP4, datacentre — the quantised build is documented on a single B200 node, and M​eta's SGLang reference configuration for it runs across four GB300 GPUs.

• Desktop, quantised — Hugging Face and the llama.cpp team tested a 2-bit quantisation on a 512GB M3 Ultra: 304GB of memory consumed, running at 18 tokens per second with room left for the KV cache. That is slow, and it is a different model from the BF16 checkpoint M​eta evaluated.

• The window you are promised and the window you can afford are two different budgets — at full 1M-token context in BF16, the KV cache alone runs to roughly 266GB. The memory figure everyone quotes is the price of loading the model, not the price of using the context length the model advertises.

Where this leaves someone who just needs a model today

Muse Spark Open is not a model you point an API key at. It is a download, a licence, a gated request form and a hardware decision, and if your workload needs an answer this week, that is a lot of surface area before the first token. The practical version of the question is which M​use model you can reach right now: Muse Spark 1.2 and Muse Spark 1.1 are the two members of the closed Muse Spark line that OrcaRouter routes, both at M​eta's list price with 0% markup, because the provider's list price is passed through rather than marked up. Muse Spark Open is not on our catalogue and we are not going to imply otherwise — a model whose entire proposition is that you can download it is a model you run yourself.

That split is also the argument for keeping both in one place. One API across 200+ models means an unproven model 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 code change. For a model released a day ago with no independent benchmark suite, that is the cheapest way to find out what it does on your traffic.

What to watch, and what would change this piece

Three things are unresolved rather than unknown, and each one would move the picture.

The first is the evaluation report M​eta says is coming. Until it lands, the model's safety claims rest on a threshold statement rather than a document.

The second is whether independent benchmarking broadens past one index. A single analyst score with no reproduction, no throughput comparison against the hosted M​use line and no audio or video measurement is a starting point, not a verdict.

The third is whether the fine-tuning ecosystem materialises. M​eta shipped TorchTitan for exactly this, and open weights earn their keep in the derivatives — merges, quantisations, domain tunes — not in the download count. Four weeks from now, the derivative trail is the honest measure of whether this release mattered.

For now the accurate summary is narrow: M​eta has published the largest US open-weight model to date, documented it thoroughly, priced it nowhere, licensed it with a revenue clause, and left the measurements to whoever does them next.

Benchmarks age and prices move; the failure record does not: the documented AI incidents and agent failures behind 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.