Meta Releases Muse Spark Open, an 816B Model, With a $20 Million Revenue Trigger in the License
A.I. / news
Meta Releases Muse Spark Open, an 816B Model, With a $20 Million Revenue Trigger in the License
The weights are on Hugging Face behind a sign-up form. Meta's own benchmark table has it trailing rivals on coding, and one outlet misreports the licence threshold.
Meta Superintelligence Labs released the weights of Muse Spark Open on October 15, an 816-billion-parameter mixture-of-experts model that activates 42 billion parameters per token, according to its Hugging Face model card. The download is gated and the licence is custom, so it is open weights, not open source.
As of Thursday the repository listed 644 likes and 2,891 downloads. Independent scores place the model in the upper tier of open-weight models on general measures and in the lower half on coding agents.
What Meta published for Muse Spark Open
The card describes 62 layers and 256 experts with 8 active per token. Context is 1 million tokens at inference. Input can be text, image, video or audio. Output is text only.
Meta's card lists NVFP4 weights as needing at least 700GB of VRAM and BF16 as needing about 2TB. It also says the model can run on two Mac Studios with M5 Ultra chips using MLX. The card gives no tokens-per-second figure for that setup.
Training data is described as public sources, web crawls, third-party data, Meta products and synthetic data. The card does not give a token count.
The licence: a 12-month, $20 million trigger
The Muse Model License permits use, modification, fine-tuning and distribution. Products must prominently display "Built with Muse Spark Open".
Companies that sell access to the model through an API and exceed $20 million in aggregate gross revenue in any 12-month period must notify Meta and sign a separate agreement first, per the card. Internal use, research and powering features of one's own product are excluded.
AI Weekly describes the threshold as $20 million per month. The model card says 12 months. This article follows the card.
Access requires a login and a form asking for full legal name, date of birth and organisation name. The card says the form cannot be edited after submission.
Scores: Meta's table, then two outside testers
The Meta-reported figures come from its own card at the highest reasoning setting. Artificial Analysis and Vals AI are independent of Meta.
| Source | Measure | Result |
|---|---|---|
| Meta card | Terminal-Bench 2.1 | 76.4 |
| Meta card | OSWorld Verified | 80.4 |
| Artificial Analysis | Intelligence Index | 39, 11th of 121 |
| Vals AI | Vals Index | 48.57%, 21st of 24 |
| Vals AI | Terminal-Bench 4.0 | 6.57%, 42nd of 52 |
Terminal-Bench 2.1 and 4.0 are different versions, so the 76.4 and 6.57 figures are not comparable. Vals AI's best result for the model is HLAB-Vals-1.1, where it ranks 2nd of 25.
Nathan Lambert called it "the strongest American-made open model by quite some margin", as quoted by AI Weekly, which did not link the original post. Lambert placed it near DeepSeek V4.1 Flash and behind GLM 5.3 and Kimi K3.
Price against other open-weight models
Artificial Analysis lists $1.25 per million input tokens and $4.25 per million output tokens, medians across six API providers. It describes that as the high end for open-weight models.
- Muse Spark Open4.25 USD
- Open-weights median1.59 USD
Source: Artificial Analysis, accessed 2026-10-22
The same page lists 140.3 output tokens per second against a 67.7 median. Meta's card says the open-weights release is not meant to be used without system-level guardrails.
Meta also released two closed models the same day, Muse Wonder and Muse Wonder Full, per AI Weekly. Agent scores shift with the harness around the model, a point the site made in its RuneBench report, where one aggregator's ranking mixed model quality with strategy choices.
The weights are live on Hugging Face now. The card does not name a date for a technical report.
Sources
More in A.I.
- 01OpenAI Says AI Found the PHP Flaws Before Attackers. Cloudflare Says They Leaked and Drew Attack TrafficThe two Oct. 21 accounts of the php2shell disclosure measure different things. Trail of Bits' Patch the Planet dashboard shows 466 of 1,995 reported issues patched.
- 02Google Cloud Lists Three OpenAI GPT-6 Models in Private Preview, With No Price and No Release DateCodex CLI can point at Google's platform today. Zero Data Retention and Confidential Compute are still listed as not yet available.
- 03Flapping Airplanes Raises $1 Billion at a Reported $6 Billion Valuation Nine Months After Its SeedThe lab's pitch is models that need far less training data. Its only technical detail so far is a GPU megakernel meant to keep chips busy.
- 04Trail of Bits Finds 21 Regressions in 130-Plus Merged Patches With a Codex Plugin, and None Touched SecurityThe security firm's regression-hunt plugin cut false positives 25% against a baseline Codex prompt, while the Patch the Planet dashboard shows 466 of 1,995 reported issues patched upstream.