Meta's Muse Spark Open Weights Come With a $20 Million Resale Threshold and a Required Credit Line
A.I. / news
Meta's Muse Spark Open Weights Come With a $20 Million Resale Threshold and a Required Credit Line
The 816B-parameter model is downloadable on Hugging Face after a gated form, and Meta's own scorecard has it behind Kimi K3 on coding and agent tests.
Meta released the weights of Muse Spark Open, an 816-billion-parameter model, under a custom Muse Model License that allows commercial use but requires an attribution line and a separate contract for large resellers. The model card on Hugging Face lists the developer as Meta Superintelligence Labs.
Trending Topics reported on Oct. 15 that the weights went public that day with a launch post and a scorecard. Meta calls the model "the most capable US open-weight model to date," the outlet reported. That is Meta's claim, and the card's own table does not fully support it.
What the Muse Spark Open licence requires
The licence permits use, modification, fine-tuning, distillation, distribution and sale. Meta claims no ownership of outputs.
Three conditions apply, according to the model card and Trending Topics:
- Products built on the model must prominently display "Built with Muse Spark Open."
- A company that sells model access as a service must notify Meta and sign a separate written agreement before offering it, once its revenue passes $20 million in any 12-month period.
- An acceptable use policy bars weapons development, malware, fraud and use by anyone under 18.
The revenue threshold does not apply to internal use, research or embedding the model in a product, Trending Topics said. Large cloud providers that want to resell it need Meta's approval. The New Stack, as cited by Trending Topics, said the terms closely resemble the licence Moonshot AI used for Kimi K3.
Specs and access
The card gives 816B total parameters with 42B active per token. The page metadata lists 822B. It is a mixture-of-experts design with 256 experts and 8 active per token, and it takes text, image, video and audio in and returns text.
Context is 1 million tokens at inference. Meta says training data includes public, crawled, third-party, Meta-derived and synthetic material, and it publishes no breakdown.
The weights are gated. Access requires a Hugging Face login and a form asking for a full legal name, date of birth and organisation name, which cannot be edited after submission.
On hardware, the card says the model "requires a GPU with at least 700GB of VRAM for NVFP4 and 2TB for BF16." It also says the model is optimised to run on two Mac Studio machines with M5 Ultra chips using MLX. The Terminal reported decode speeds for that setup earlier.
What Meta's scorecard shows
Every figure below is Meta-reported, run at its highest reasoning setting, with no independent reproduction. The card compares against Kimi K3 and GLM 5.3.
| Benchmark | Muse Spark Open | Kimi K3 |
|---|---|---|
| Terminal-Bench 2.1 | 76.4 | 85.0 |
| OSWorld Verified | 80.4 | 84.8 |
| DeepSWE 1.1 | 60 | 68 |
| BabyVision | 80.2 | 85.7 |
- Kimi K385 points
- Muse Spark Open76.4 points
Source: Muse Spark Open model card on Hugging Face, accessed 2026-10-18
Muse Spark Open leads the comparison on one test, MCP-Atlas, at 89.2. The card gives its Artificial Analysis Intelligence Index score as 39, against 45 for GLM 5.3 and 44 for Kimi K3. The model had not yet been listed in Artificial Analysis's open-weight ranking when Trending Topics ran its piece.
The card also reports safety numbers that Meta measured itself: 99.2% refusal on its internal cyber jailbreak set and 98.1% on the MASK honesty test.
What is still unknown
Meta has not said how many tokens the model trained on or what share was synthetic. The model card names no independent evaluator. Another Hugging Face release this month, DeepSeek's V4.1-Flash-1017, ships under the MIT licence, which carries no resale clause.
Sources
More in A.I.
- 01Vals Test: Three of Six AI Agents Threatened Court Over a £350 Train RefundTold to try everything to recover a refund the rules no longer allowed, Grok 4.7 sent 187 messages in 15 simulated days while Opus 5.5 sent eight in twelve weeks.
- 02MiniCPM-V 4.7 Is Apache 2.0 With 3B Active Parameters and No Benchmark TablesOpenBMB's first mixture-of-experts vision model claims a win over Qwen3.6 35B-A3B in prose and charts, and 13 tokens a second on an iPhone, but the model card prints no scores.
- 03Gemini 4 Argon Opens to the API at $2 per Million Tokens, and Loses 5 of 19 Rows in Google's Own TableGoogle's October 13 launch post leads with a Vals Index ranking; the comparison table behind it shows Argon behind GPT-6 Astra or Opus 5.5 on both Terminal-Bench tests and on OSWorld 2.0.
- 04StarSkirmish Has Nine Models Write StarCraft Bots in One Hour, and the Best Wins 7% Against Top Human BotsGemini 4 Argon lands in a three-way tie at 37 to 39, but it ran at a lower reasoning setting than most rivals and its runs vary the most.