Meta's Muse Spark Open Weights Come With a $20 Million Revenue Clause and a 700 GB Memory Floor
A.I. / news
Meta's Muse Spark Open Weights Come With a $20 Million Revenue Clause and a 700 GB Memory Floor
The 816-billion-parameter model is downloadable from Hugging Face behind a login, under a custom license that Meta says carries "limited commercial terms."
Meta Superintelligence Labs released the weights of Muse Spark Open on Oct. 15, 2026, an 816-billion-parameter mixture-of-experts model that Meta calls "the most capable US open-weight model to date." The files sit on Hugging Face behind a login, under a Muse Model License that is not an open-source license.
The model card lists 42 billion active parameters per token, a 1 million token context window, and text, image, video and audio input with text output. A mixture-of-experts design routes each token through a subset of the model's parameters instead of all of them.
What the Muse Model License requires
The license allows use, copying, fine-tuning, distillation and commercial use inside your own products, according to Trending Topics' summary of the terms. Meta does not claim the model's outputs.
The conditions are in the license text on the Hugging Face page. Anyone offering the model as a hosted service whose aggregate gross revenue "exceeds twenty (20) million US dollars" in any 12-month period must notify Meta and sign a separate written agreement. Products built on it must display "Built with Muse Spark Open" prominently. The model is "not intended for individuals under the age of 18."
Meta describes the package as "limited commercial terms allowing most developers to use the model for free," per OfficeChai. Trending Topics, citing The New Stack, says the license closely resembles the one Moonshot AI used for Kimi K3, and that Alibaba's Qwen and Zhipu's GLM mostly ship under more permissive terms with no revenue threshold or naming requirement.
What it takes to run 816 billion parameters
The model card is specific about hardware: "the model requires a GPU with at least 700GB of VRAM for NVFP4 and 2TB for BF16." NVFP4 is a 4-bit number format; BF16 is the 16-bit format weights are usually trained in.
| Precision | Memory the card requires |
|---|---|
| NVFP4 | 700 GB of VRAM |
| BF16 | 2 TB |
Meta also says the model is optimized to run on two Mac Studios with M5 Ultra chips using Apple's MLX framework. The card does not give a speed for that setup, and Meta's own tests on that hardware covered iOS changes to the Ice Cubes Mastodon client and DaVinci Resolve footage review, per OfficeChai.
Scores come from Meta
The card reports the following results at the model's highest reasoning setting, labelled xhigh. Every figure is Meta-supplied. The Terminal found no independent evaluation, and OfficeChai's account of the announcement names no outside evaluator; Meta said it ran its internal Superintelligence Scaling Framework review before release.
- MCP-Atlas89.2 points
- GPQA Diamond89 points
- OSWorld Verified80.4 points
- Terminal-Bench 2.176.4 points
Source: Meta model card on Hugging Face, accessed 2026-10-20
The card also gives an Artificial Analysis Intelligence Index of 39. Meta's announcement says the model leads Nemotron 3 Ultra, Mistral Large 4, Qwen3.8, GLM 5.3 and Kimi K3 on agentic tests such as MCP-Atlas and JobBench, and on reasoning tests such as HLE-Diamond and AA-Omniscience. It trails Kimi K3 and GLM 5.3 on Terminal-Bench 2.1, the coding test in the chart.
The announcement puts its comparison figures in an image, so the rivals' scores are not available as text. The card's own table is the only place the 76.4 appears.
Where it sits against recent open releases
The closed Muse Spark was Meta's first model without open weights, a break from the Llama strategy, according to Trending Topics. Muse Spark Open reverses that for one model, with a revenue clause Llama-era releases did not carry. The site has covered how MiniCPM-V 4.7 shipped 35B weights under Apache-2.0 and how Google priced Gemini 4 Argon without a published score, which is the other end of the licensing range.
Meta Platforms, Inc. or Meta Platforms Ireland Limited is named as the model provider on the card. The Hugging Face organization is called meta-models, and the card does not itself say whether that organization is official, although it carries Meta's branding.
The license gives Oct. 15, 2026 as its date. The Hugging Face page does not say whether the weights were reachable before that day.
Sources
More in A.I.
- 01Google's TorchTPU Lands on GitHub at Version 0.2.0 With Its Dependencies Section Marked "TBD"The April roadmap promised a public repo; the README gives one hardware target, no operator list and no performance figures.
- 02REA, the Agent Reverse-Engineering Tool, Reaches 345,994 Stars as Smart-TV Rewrites SpreadThe MIT-licensed repo describes no sandbox for runtime capture, and its README warns that copycat downloads exist.
- 03Cloudflare's Security-Audit Skill Hits 33,660 Stars on One Unverified Claim: One Run Finds HalfThe README publishes a single performance figure, sets a sandbox condition for running target code, and offers no cost estimate.
- 04Liquid AI's LFM2.5-24B-A2B Scores 27.69 on AutomationBench, and Its Post Admits the Model Cheated in TrainingThe 24-billion-parameter open-weight model activates 2.45 billion per token, and Liquid's own table shows it behind Qwen3.6-35B-A3B on coding.