Meta’s newest model arrived last Thursday with two headline numbers and one asterisk. The numbers: 816 billion parameters and open weights. The asterisk: the license. Muse Spark Open, which Meta calls the most capable US open-weight model to date, can be run, tuned, and redistributed by almost anyone, but its terms add a revenue trigger, a branding rule, and a usage policy that shuts out some customers. Here is what comes with Muse Spark Open, how its scores stack up against Chinese rivals, and what the fine print means for anyone building on it.
What Meta released with Muse Spark Open
Muse Spark Open has 816 billion parameters in total, but only 42 billion do the work on any given token, a mixture-of-experts setup described in The New Stack’s launch report. It reads up to 1 million tokens of context. Let’s Data Science adds that each token is routed through 8 of 256 experts, and that the model takes text, image, video, and audio as input and answers in text.
Developers can pull the weights from Amazon SageMaker, Google Cloud Model Garden, Hugging Face, or the Microsoft Foundry catalog. Anyone who would rather rent it has hosted APIs at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, where Together charges $1.25 per million input tokens and $4.25 per million output tokens, per Let’s Data Science. Meta shipped a closed version of Spark in April, and this is its open sibling.
The license, in plain terms
Meta describes the Muse Model License as having limited commercial terms. Weights under it can be run, modified, fine-tuned, distilled, and redistributed, and Meta claims no ownership of outputs. Internal use and building the model into your own product cost nothing. Anything built on Muse Spark Open has to show the phrase “Built with Muse Spark Open” in its interface.
Companies that sell the model itself as a hosted service face the real condition. Once a company and its affiliates pass $20 million in gross revenue over any 12 months, they have to notify Meta and sign a separate written agreement first. The usage policy also rules out military and critical-infrastructure deployments, with a carve-out for the US government, the same pattern Meta used for Llama in 2024. The New Stack’s author called the result a semi-open license: open weights, but terms that are not open in the Apache sense.
Where it leads and where it trails
Meta’s own table puts Muse Spark Open ahead of Chinese open models on agentic work. On MCP-Atlas, a tool-use test, it scored 89.2 against 84.5 for Qwen3.8, 84.2 for GLM 5.3, and 82.3 for Kimi K3. On JobBench it scored 61.8, just above GLM 5.3 at 61.4. Meta also says it beats Thinking Machines’ Inkling and Nvidia’s Nemotron 3 Ultra across the board, often by 20 points or more.
These tests measure different jobs. Tool-use tests ask a model to pick the right function and call it correctly across many steps, while coding tests ask it to work in a terminal or fix real software repositories. A model can lead one and trail the other, and that is what happened here.
Coding is the weak spot. On Terminal-Bench 2.1, Muse Spark Open scored 76.4 against 85.0 for Kimi K3 and 83.9 for GLM 5.3. On DeepSWE 1.1 it scored 60, against 68 for Kimi K3 and 67 for GLM 5.3. Meta acknowledged the same gap when it launched the closed model. The overall Artificial Analysis index figure in Meta’s table, 39, is the lowest of the four models, and Let’s Data Science notes it has no independent confirmation yet.
Running it at home
Meta also pitched a desktop option. The company says two Mac Studios with M5 Ultra chips and 512GB of memory each can be linked to run the model using Apple’s MLX framework. Mark Zuckerberg posted that the weights run great on the new machines. Compressed weights take about 493GB, which is why one machine is not enough. The catch is availability: the 512GB configuration is not on sale yet, and Let’s Data Science says it ships in late October with no price announced.
Four 256GB Mac Studios are available now at $10,799 each, or $43,196 before tax. Apple’s benchmark showed more than 36 tokens per second on four Macs at an 8,192-token context. Those are Apple’s figures, not independent measurements. For most small teams, a hosted API at those token prices will cost less than the hardware.
What it means for developers and companies
A startup that wants to build an agent on Muse Spark Open can start today at no license cost, provided it shows the attribution line and stays out of the barred uses. A hosting company is in a different position, since the $20 million revenue line turns a free download into a negotiated deal. Military deployments are barred, with the carve-out for the US government.
The release also lands in a crowded field. Chinese open models are the yardstick Meta chose, and we looked at the hardware side of that race in our piece on Nvidia’s China AI chip share. Meta is spending heavily on AI, as we described in our preview of its earnings and AI spending. Muse Spark Open looks like a bet that tool use, not coding, will decide which open model companies pick.
What to watch next
Three things will settle how good it is. First, independent results from Artificial Analysis and others, since nearly every number above comes from Meta. Second, whether coding scores improve in a later release. Third, the actual license behavior of hosting providers, because their agreements with Meta will show how the revenue clause works in practice. Nathan Lambert of Interconnects wrote that he had waited a long time for US open-weight models to matter again, and added that there is still more work to do.
Editor’s take
Muse Spark Open is a strong release with a conditional welcome. It leads on the agentic tests that companies care about and gives US developers an open-weight option, but it lags Chinese models on coding, and its terms are more restrictive than the word open suggests. Treat Meta’s benchmark table as a claim until independent tests arrive. For now, a developer should read the license before the leaderboard.
