Fine-tune on Muse Spark Open outputs and you owe Meta a credit line

The first paragraph reads like the MIT licence, word for word in places. Meta put the weights of Muse Spark Open on Hugging Face on 15 October: 816B parameters in total, 42B active, a 1M-token context, and a licence that starts "free of charge" and "without restriction". Keep reading, though. Two pages down, the definition of a derivative reaches every model you train on its outputs, and that's the part you'll want to read before you point it at a synthetic data job.

The clause that reaches past the weights

Section 2 of the licence asks you to "prominently display" the line "Built with Muse Spark Open" in the interface of every product or service that uses the model or a Derivative, or in its documentation, website or settings menu. That's a light ask. A line in an About page covers it.

The weight sits in the definition. A Derivative is "any AI model created, trained, modified (including fine-tuned), distilled, or otherwise improved using the Software or any outputs or results thereof." Outputs. So if you generate a training set with Spark Open and fine-tune a small Qwen on it, your small Qwen is a Derivative under this text, and Section 1 says the licence and the Usage Policy have to travel with copies of Derivatives too.

Here's the odd bit. The model card lists "Synthetic Data Generation" and "LLM-as-a-judge Evaluation" among the things the model does well. Both are output-to-model pipelines. We read "otherwise improved using... results thereof" as broad enough to catch a model you tuned against Spark Open's judge scores, though I might be wrong about how Meta would read it, and the licence doesn't give an example. If your team runs a distillation pipeline, this isn't a reason to skip the model. It's a reason to decide up front which of your models will carry the credit line, and to keep Spark Open out of pipelines whose outputs you plan to ship under a different licence.

Who has to call Meta first

Section 3 is narrower than the headlines make it sound. It only bites when you or your affiliates sell access to AI models through an API or a similar interface, which the licence calls Model as a Service, and the group's gross revenue passes $20 million over any 12 consecutive months. Then you notify Meta and sign a separate agreement before offering Spark Open or any Derivative that way. The licence spells out two exemptions: internal use, research, or any commercial work that isn't Model as a Service, and using the model to power features of your own product. A $500 million company running it behind its own help desk doesn't need that agreement. A $25 million inference host does.

Then there's the Usage Policy, attached as Exhibit A and binding through Section 1. Most of it is the list you'd expect. One line deserves a second look from the ops crowd: you shouldn't use the model for the "operation of critical infrastructure, transportation technologies, or heavy machinery". An agent that reads alerts and suggests fixes is probably fine. One that acts on a utility's control plane is the case that line was written for. The policy also says the model isn't meant for anyone under 18.

Access isn't instant either. Both repos are gated with manual review, and the form asks for your legal name, date of birth, country, affiliation and job title, and warns that you can't edit it after you submit. Compare Muse Glimmer, Meta's 30B model from August: Apache 2.0 and ungated. Same company, two months apart, very different paperwork.

What you need to run it, and where it lands

We added up the file sizes from the Hugging Face API on 16 October. The BF16 repo holds 1,644 GB, the NVFP4 repo 533 GB. The card asks for "at least 700GB of VRAM for NVFP4 and 2TB for BF16", which leaves room for the KV cache on top of the weights. Meta's developer post says the NVFP4 build runs on two Mac Studios with M5 Ultra and 512 GB each, or four with 256 GB, linked over Thunderbolt 5. The card names SGLang, vLLM, TensorRT-LLM, Transformers, Unsloth and llama.cpp as supported. One default to watch: the chat template sets reasoning effort to xhigh unless you pass low, medium or high, so a quick test can run longer than you'd expect.

On quality, read Meta's own table rather than the "most capable US open-weight model" line. Spark Open leads on JobBench (61.8) and MCP-Atlas (89.2), which are agentic tool-use tests. On the Artificial Analysis Intelligence Index it scores 39, against 45 for GLM 5.3 and 44 for Kimi K3. On Terminal-Bench 2.1 it's 76.4 to Kimi K3's 85.0, and on DeepSWE 1.1 it's 60 to 68. That's Meta's table, run with Meta's settings, and it still has Spark Open trailing the two Chinese models on most coding rows. The "US" in the tagline is doing real work.

Our take: for internal agents and tool-calling work it's worth a test, and the licence is easy to live with there. For distillation or a resale API, read the whole licence with your lawyer first. We've added it to the open-weight models tracker.

FAQ

Sources: Meta, Introducing Muse Spark Open and Build with Muse Spark Open (15 October 2026); the Hugging Face model card and Muse Model License; the NVFP4 repo; independent: Hugging Face's launch post and Unite.AI. File sizes read from the Hugging Face API on 16 October 2026.