Muse Spark Open Is Free. Every Host Charges $1.25 Anyway.

Meta released the weights for Muse Spark Open on 15 October 2026: 816B total parameters, 42B active, natively multimodal, a 1M-token context window, and a name it chose itself — “the most capable US open-weight model to date.” Built for coding and agentic work, it is downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, and served day one by seven API platforms: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

Three numbers make the “free” more complicated than the announcement’s sentence suggests. All five hosts that publish prices list the identical $1.25 / $4.25 per million tokens — open weights have not moved the hosted price at all. The license grants MIT-style freedom and then adds two conditions, one of which has no revenue threshold at all. And the launch demo of the model running on your desktop ran on two Mac Studios whose 512 GB configuration — Apple’s own words — arrives “in late October.”

Five hosts, one price

From OpenRouter’s provider table, per million tokens, checked 19–20 October:

Four independent trackers of provider price lists (apipriceindex, ctrlaltdebrief, llmhosting.ai, llmreference) agree with OpenRouter’s table, and add that Groq — the third platform named in Meta’s own announcement — shows “Contact sales” instead of a rate. When every seller of the same free weights prints the same number, price stops being a differentiator: you choose a host on latency, throughput and cache hit rate. OpenRouter’s own effective-price table says what caching is worth — a weighted average of $0.78/M on input against the $1.25 list, with the best providers at $0.38–$0.45 on traffic that hits cache 72–79% of the time.

For context on what the number itself means: on the public price sheets, Sonnet 5.5 is $2 / $10, Meta’s own closed Muse Wonder is $4 / $20, and Claude Haiku 5.5 is $0.10 / $0.50 below its 100,000-token cliff. The open model priced like a mid-tier closed one is not a price story. It is a capability and license story.

The license: one badge, one $20 million line

The Muse Model License on the Hugging Face model card opens with an MIT-style grant — use, copy, modify, fine-tune, distill, merge, publish, distribute, sublicense, sell — and then gets specific.

Clause 2 is a branding requirement with no revenue exemption:

“Licensee shall prominently display ‘Built with Muse Spark Open’ in the user interface of each product or service that uses the Software or any Derivatives, or in the accompanying documentation, website, or settings menu.”

“Derivatives” is defined broadly: any model created, trained, modified, fine-tuned, distilled “or otherwise improved using the Software or any outputs or results thereof.” Fine-tuning on your own data distilled from this model’s outputs is, by that text, a Derivative carrying the badge.

Clause 3 is the revenue line, and it is narrower than it first reads:

“If the Licensee or any of its affiliates provide a third party with access to artificial intelligence models as a product or service, via API or a similar interface … and the aggregate gross revenue of the Licensee and its affiliates exceeds twenty (20) million US dollars … during any consecutive twelve (12) month period, the Licensee must notify Meta and enter into a separate written agreement with Meta before offering the Software or any Derivatives as part of that Model as a Service business.”

The carve-outs do the work: internal use, research, and “the Licensee’s use of the Software as a component to power features of the Licensee’s own product or service that is not a Model as a Service” are outside it. So the clause targets companies that resell model access for a living, not SaaS companies that embed the model. For shape: Llama 3’s community license set its famous threshold on monthly active users; this one is revenue-based, applies only to Model-as-a-Service sellers, and adds the always-on badge. Meta’s own August release, the 30B Glimmer distilled from Muse Spark, reportedly went out under Apache 2.0 — smaller models free of conditions, the flagship’s weights with two.

Clause 4 gives Meta no ownership of outputs. Clause 5 disclaims all warranties and says Meta has no liability for derivatives — the licensee is “solely responsible for the safety, deployment, and use” of them.

The desktop demo runs on hardware that was not shipping

Meta’s framing for local use: Muse Spark Open is “the first model in its class to run on your desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework,” and “two Mac Studios behave like one supercomputer with a terabyte of unified memory.”

The hardware numbers are Apple’s: the M5 Ultra Mac Studio starts at $5,499 with 96 GB of unified memory; the 256 GB configuration Engadget priced runs $11,299 with 2 TB of storage; and the 512 GB configuration the claim depends on — 1.2 TB/s each, a terabyte as a pair — was announced on 25 August as “coming in late October.” At release day, the demo machine was on order, not on a shelf.

Which quantization that terabyte needs is arithmetic on the Hugging Face listing: the BF16 checkpoint carries 822B parameters — roughly 1.6 TB of weights at 2 bytes each, more than two 512 GB machines hold, before any KV cache. The NVFP4 checkpoint carries 484B parameters, which fits. So the honest reading of “frontier on your desktop” is: a pair of $5,499-and-up machines in a quantization you did not pick, at a 512 GB price Apple has not published — a five-figure entry for weights that cost nothing, which is the same trade Kimi K3 made on a marketplace from the other direction: one model, and the hosting decision still has to be made by hand.

What it beats, and the coding gap Meta published itself

From Meta’s model card and launch table, as mirrored by benchmarklist and The New Stack:

Meta says Spark Open beats the two US rivals in its class — Thinking Machines’ Inkling and Nvidia’s Nemotron 3 Ultra — “often by 20 points or more” on its own slide. Against the Chinese open-weight incumbents the picture splits: it leads on agentic tool use, ties on scientific coding, and loses the plain coding benchmarks — the same gap The New Stack notes Meta admitted when it shipped the closed Muse Spark in April. Alexandr Wang, Meta’s chief AI officer, put the strategy plainly in launch-week coverage: Muse Spark Open is deliberately less capable than Monday’s closed Muse Wonder, and Meta “will generally not open-source our most powerful models.”

What we could not verify

• Clause 1 of the license. Our fetch of the model card returned clauses 2–5 and the grant in full; clause 1 — in an MIT-style grant typically the notice sentence — was cut. We therefore cannot say whether a usage policy or other condition sits outside what we quoted.
• Whether any of the five hosts signed clause-3 agreements. They sell Model as a Service; whether they are over the $20M threshold and what they signed with Meta is not public.
• Groq’s price may simply not be published yet, rather than absent.
• The 512 GB Mac Studio’s price and shipping status today. “Late October” was Apple’s 25 August window; we did not re-check the store.
• The 256k maximum output appears on one tracker (apipriceindex); we did not read it from Meta or a host.
• We ran no benchmarks and no local runs. Every number above is Meta’s, Artificial Analysis’s, or aggregators’ mirrors of those; benchmarklist marks much of it “Self-reported.”

If you are the one buying

1. Buying API access: pick among the five by latency and throughput, not price — it is one number at every host. Then cache: $0.15 cache reads turn a $1.25 sticker into $0.38–$0.45 effective on cache-heavy traffic.
2. Building a product on it: put the badge on your compliance list on day one. Clause 2 applies from your first dollar, “settings menu” is allowed, and fine-tunes distilled from the model’s own outputs inherit the condition.
3. Selling model access, group revenue over $20M: clause 3 means a written agreement with Meta before you launch the feature — budget the legal time, not just the tokens.
4. Chasing the desktop claim: wait for the 512 GB price list, plan around five figures, and treat the quantization as Meta chose it. The $5,499 machine is the entry to the conversation, not the answer.
5. Choosing a coding agent: open is not a reason. On the coding benchmarks Meta itself published, Kimi K3 scores 85.0 to Spark Open’s 76.4 — and hosts it for a fraction of $1.25. Buy Spark Open for the tool-use and agentic boards, where it currently leads.