Muse Spark Open's license asks hosted-model sellers above $20M in revenue for a written deal with Meta
Meta released Muse Spark Open, an 816B-parameter open-weight model that activates 42B parameters per token. Per The New Stack, a company above $20M in revenue needs a separate Meta agreement before selling it as a hosted service.
The Engineer · Build desk
Sellers of the model as a service above $20 million in gross revenue need a written deal with Meta. API hosts such as Groq serve it. Companies using it must display 'Built with Muse Spark Open'. Internal and own-product use is free; military and critical infrastructure use is barred.
- What it limits constraint Model-as-a-service sellers Above $20 million gross revenue: notify Meta and sign a separate written agreement before offering it as a service, claim 5
- Who's exposed exposure API hosts Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI serve the model through their APIs, claim 2
- What it limits constraint Any company using the model Must prominently display 'Built with Muse Spark Open' in each product using it or a derivative, claim 7
- What's now possible capability Internal users, product builders Internal use and building the model into your own product are free, per The New Stack, claim 8
- What it limits constraint Military, infrastructure users The usage policy bars military and critical infrastructure deployments, per The New Stack, claim 8
| Who | How | Kind | Claim |
|---|---|---|---|
| Model-as-a-service sellers | Above $20 million gross revenue: notify Meta and sign a separate written agreement before offering it as a service | constraint | 5 |
| API hosts | Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI serve the model through their APIs | exposure | 2 |
| Any company using the model | Must prominently display 'Built with Muse Spark Open' in each product using it or a derivative | constraint | 7 |
| Internal users, product builders | Internal use and building the model into your own product are free, per The New Stack | capability | 8 |
| Military, infrastructure users | The usage policy bars military and critical infrastructure deployments, per The New Stack | constraint | 8 |
What happened
- Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI each serve the model through their APIs.
- Any company using the model must prominently display 'Built with Muse Spark Open' in products that use it or a derivative trained, fine-tuned or distilled from it.
- Internal use and building the model into your own product are free, and the usage policy bars military and critical infrastructure deployments.
- The New Stack reports it leads Kimi K3, GLM 5.3 and Qwen3.8 on agentic and tool-use benchmarks such as JobBench and MCP-Atlas.
Why it matters
- Who has to decide decision The first license question is whether you sell the model as a service. On The New Stack's account, a team above the line that builds it into its own product owes the attribution line and must stay within the usage policy.
- What doesn't add up contradiction Meta pitches Muse Spark Open as a coding model, yet coding is where the reported scores trail Kimi K3 and GLM 5.3. A coding-heavy workload is where Meta's numbers transfer least.
- Who's exposed exposure The seven API hosts sell the model as a service, so any of them above the revenue line falls under the clause. The sources do not say what agreements they hold, so buyers will have to ask each host.
- What it sets up precedent On The New Stack's comparison, Meta's terms resemble Moonshot's for Kimi K3. A license read now has to be repeated for each strong open-weight model a team compares.
The exact wording decides who has to call Meta. A company over the revenue line 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," according to The New Stack's quotation of the license. [5] We read "Derivatives" as reaching past the base weights to variants built on them. Meta's own summary is that the license has "limited commercial terms, which allow many developers to use the model for free." [6] The New Stack says it looks a lot like the license Moonshot uses for Kimi K3. [9]
Meta's lead over US peers comes from Meta's own tests. The model beats Thinking Machines' Inkling and Nvidia's Nemotron 3 Ultra across the board, often by 20 points or more. [24] Against the Chinese models the picture splits. The New Stack reports a rough tie on SciCode [10] and a coding deficit: 8.6 points behind Kimi K3 and 7.5 behind GLM 5.3 on Terminal-Bench 2.1 [20], plus an 8-point spread on DeepSWE 1.1 [21].
That does not fit the pitch. Meta describes a coding-optimized model trained on real developer workflows such as multi-file refactors and long debugging sessions. [4] The New Stack adds that Meta acknowledged the same gap when it launched the closed version of Spark in April. [13] For the agentic numbers to transfer, your workload has to resemble the tool-use tasks where the model leads, not the Terminal-Bench and DeepSWE tasks where it trails.
Self-hosting has a different constraint: memory. The weights are on four platforms: Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft's Foundry catalog. [3][19] Each MoE block has 256 experts, and a router selects 8 for each token. [14] That puts about 42B parameters, roughly 5.1% of the 816B, to work on each token. [23]
All 256 experts still sit in memory. Hugging Face says deployment therefore needs the memory of an 816B model and runs at the speed of a 42B one. [15] The BF16 checkpoint is about 1.5 TB, and in transformers it runs across 4 nodes of 8 H200 GPUs, 32 in all, with tensor parallelism of 32. [16][22] The 4-bit NVFP4 checkpoint targets Blackwell GPUs and runs on a single B200 node, and Hugging Face reports negligible accuracy loss. [16][17]
What to watch
- The full text of the Muse Model License and usage policy, especially how they define Model as a Service and Derivatives for fine-tuned internal tools.
- Whether the seven API hosts disclose the terms they hold with Meta.
- Independent runs of Terminal-Bench 2.1 and DeepSWE 1.1 on the released weights, to see if the coding gap holds outside Meta's setup.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption40
- Hype gap+20
- Incentives60
- Confidence70
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Meta launched Muse Spark Open, a model with 816B parameters and 42B active parameters, which the company argues is the 'most capable US open-weight model to date.'
- [2]
Muse Spark Open is available through the APIs of Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
- [3]
The weights can be downloaded from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft's Foundry Model Catalog.
- [4]
Meta describes Muse Spark Open as a coding-optimized model trained on real developer workflows such as multi-file refactors and long debugging sessions.
- [5]
Under the Muse Model License, any company with more than $20 million in gross revenue during any consecutive twelve month period 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.'
- [6]
Meta says the Muse Model License has 'limited commercial terms, which allow many developers to use the model for free.'
- [7]
Any company using the model must prominently display 'Built with Muse Spark Open' in the user interface of each product or service that uses the model or a derivative model created, trained, fine-tuned, or distilled from it.
- [8]
Internal use and building the model into your own product are free, and the usage policy that ships with the license bars military and critical infrastructure deployments.
- [9]
The New Stack says the Muse Model License looks a lot like the license Moonshot uses for Kimi K3.
- [10]
Muse Spark Open leads Kimi K3, GLM 5.3, and Qwen3.8 on agentic and tool-use benchmarks, including JobBench and MCP-Atlas, and roughly ties them on SciCode.
- [11]
On Terminal-Bench 2.1, Kimi K3 scores 85.0 and GLM 5.3 scores 83.9, against 76.4 for Muse Spark Open.
- [12]
On DeepSWE 1.1 the spread is 68 to 60.
- [13]
The coding gap is the same one Meta already acknowledged when it launched the closed version of Spark in April.
- [14]
Each MoE block in Muse Spark Open has 256 experts, and a router selects 8 experts for each token.
- [15]
All 256 experts stay in memory, so deployment needs the memory of an 816B-parameter model but runs at the much faster speed of a 42B model.
- [16]
The BF16 checkpoint is about 1.5 TB in safetensors format. In transformers it runs across 4 nodes of 8 H200 GPUs with tensor parallelism (TP=32), and the NVFP4 checkpoint runs on a single B200 node.
- [17]
The NVFP4 checkpoint stores the parameters in 4-bit precision with negligible loss in accuracy, specifically for NVIDIA Blackwell GPUs.
- [18]
Seven inference providers offer Muse Spark Open through APIs.
- [19]
The weights are available from four download platforms.
- [20]
On Terminal-Bench 2.1, Muse Spark Open trails Kimi K3 by 8.6 points and GLM 5.3 by 7.5 points.
- [21]
The DeepSWE 1.1 spread is 8 points.
- [22]
The BF16 deployment in transformers uses 32 H200 GPUs.
- [23]
About 42B of the 816B parameters, roughly 5.1%, are active on each token.
- [24]
In Meta's own benchmarks, Muse Spark Open beats Thinking Machines' Inkling and Nvidia's Nemotron 3 Ultra across the board, often by 20 points or more.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
4 independent publishers whose own reporting we read for this story.
- developer.nvidia.comScale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72
1 article · October 15, 2026
- huggingface.coWelcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta
2 articles · October 14, 2026
- pytorch.orgMuse Spark 1.2 Is Now Open Weights
1 article · October 15, 2026
- thenewstack.ioMeta ships Muse Spark Open, its answer to China’s open-weight models
1 article · October 15, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.