In short
- Meta says Muse Spark Open has 816 billion parameters, activates 42 billion per token and has a context window of one million tokens.
- Under the Muse Model License, Model as a Service providers above 20 million US dollars in gross revenue over any consecutive 12 months must sign a separate agreement.
- Meta's own table scores it 39 on the Artificial Analysis Intelligence Index, below GLM 5.3 (45) and Kimi K3 (44), and 89.2 on MCP-Atlas, above both (84.2, 82.3).
Meta released Muse Spark Open on 15 October 2026. It has 816 billion parameters in total and activates 42 billion per token, with a context window of one million tokens. It is a mixture-of-experts model with 256 experts per layer, of which 8 are used for each token.
It takes text, images, video and audio as input and returns text. The licence is the Muse Model License.
It allows use, modification, fine-tuning and distribution, requires products built on it to display "Built with Muse Spark Open", and requires a separate agreement with Meta for Model-as-a-Service offerings once the provider's aggregate gross revenue exceeds 20 million US dollars in any 12 months.
Internal use and research are excluded from that condition. Meta's model card compares it with GLM 5.3 and Kimi K3: it scores 39 on the Artificial Analysis Intelligence Index against 45 and 44, and 76.4 on Terminal-Bench 2.1 against 83.9 and 85.0, but 89.2 on MCP-Atlas against 84.2 and 82.3.
The BF16 checkpoint runs across four nodes of eight H200 GPUs; the NVFP4 version serves on four Blackwell Ultra GPUs or eight H200s. Meta's post also describes serving it on two Mac Studios with M5 Ultra and 512GB of unified memory.
What it means for you · our view
The benchmark figures come from Meta's own model card table, so treat the comparison as a vendor claim until an independent evaluation repeats it.
On that table the model trails GLM 5.3 and Kimi K3 on the Intelligence Index and Terminal-Bench 2.1 and leads on MCP-Atlas, which points to testing the measure closest to your workload. The licence matters as much as the weights.
Internal use and research sit outside the Model as a Service condition, but a team that offers access through an API should read the revenue threshold and the display requirement before building on it. Ask your legal team whether your planned use counts as Model as a Service, and run your own test set before any switch.
How we checked this · the sources15 checked facts from 4 sources.
-
Build with Muse Spark Open · Meta · archived copy
- Meta released Muse Spark Open on 15 October 2026.
checked 16 Oct 2026Oct 15, 2026 ... An 816B-parameter open-weight model from Meta Superintelligence Labs
- Muse Spark Open has 816 billion parameters in total and activates 42 billion parameters per token.
checked 16 Oct 2026An 816B-parameter open-weight model from Meta Superintelligence Labs ... It activates 42B parameters per token
- Muse Spark Open has a context window of one million tokens.
checked 16 Oct 2026holds a 1M-token context window
- Meta's post says the NVFP4 weights serve on four Blackwell Ultra GPUs or eight H200s.
checked 16 Oct 2026Serves on four Blackwell Ultra GPUs or eight H200s.
- Meta's post says the model can be served on two Mac Studios with M5 Ultra and 512GB of unified memory.
checked 16 Oct 2026Serve the model on either two Mac Studios, with M5 Ultra and 512GB of unified memory;
- Meta released Muse Spark Open on 15 October 2026.
-
Muse Spark Open model card · Meta · archived copy
- Muse Spark Open is a mixture-of-experts transformer with 256 experts, of which 8 are active per token.
checked 16 Oct 2026Mixture-of-Experts Transformer, with Perception Encoder ... Number of Experts: 256 ... Experts Activated per Token: 8 (top-k routing)
- Muse Spark Open takes text, image, video and audio as input and its output is text only.
checked 16 Oct 2026Input: Text + Image + Video + Audio / Output: Text
- On the model card's benchmark table Muse Spark Open (xhigh) scores 39 on the Artificial Analysis Intelligence Index, against 45 for GLM 5.3 (max) and 44 for Kimi K3 (max).
checked 16 Oct 2026Artificial Analysis Intelligence Index | 39 | ... | GLM 5.3 (max) 45 | Kimi K3 (max) 44
- On the model card's benchmark table Muse Spark Open (xhigh) scores 76.4 on Terminal-Bench 2.1, against 83.9 for GLM 5.3 (max) and 85.0 for Kimi K3 (max).
checked 16 Oct 2026Terminal-Bench 2.1 (Terminus 2 agent) | 76.4 | ... | GLM 5.3 (max) 83.9 | Kimi K3 (max) 85.0
- On the model card's benchmark table Muse Spark Open (xhigh) scores 89.2 on MCP-Atlas, against 84.2 for GLM 5.3 (max) and 82.3 for Kimi K3 (max).
checked 16 Oct 2026MCP-Atlas | 89.2 | ... | GLM 5.3 (max) 84.2 | Kimi K3 (max) 82.3
- Muse Spark Open is a mixture-of-experts transformer with 256 experts, of which 8 are active per token.
-
Muse Model License · Meta · archived copy
- The licence is the Muse Model License, which permits use, modification, fine-tuning and distribution.
checked 16 Oct 2026the rights to use, copy, modify (including fine-tune), create derivative works, distill, merge, publish, distribute
- The licence requires products or services using the model to prominently display "Built with Muse Spark Open".
checked 16 Oct 2026Licensee shall prominently display "Built with Muse Spark Open" in the user interface of each product or service
- The licence requires a licensee offering the model as a service (access via API or a similar interface) to notify Meta and sign a separate written agreement first, once the licensee and its affiliates exceed twenty (20) million US dollars in aggregate gross revenue during any consecutive twelve (12) month period.
Toned downchecked 16 Oct 2026the aggregate gross revenue of the Licensee and its affiliates exceeds twenty (20) million US dollars ... during any consecutive twelve (12) month period ... the Licensee must notify Meta and enter into a separate written agreement with Meta before offering the Software
- The licence's Model as a Service condition does not apply to internal use, research, or commercial activity that is not Model as a Service.
checked 16 Oct 2026This Section 3 does not apply to (a) internal use, research, or any commercial activity that is not Model as a Service
- The licence is the Muse Model License, which permits use, modification, fine-tuning and distribution.
-
Muse Spark Open on Hugging Face · Hugging Face · archived copy
- Hugging Face says the BF16 checkpoint runs across 4 nodes of 8 H200 GPUs with tensor parallelism (TP=32).
checked 16 Oct 2026
- Hugging Face says the BF16 checkpoint runs across 4 nodes of 8 H200 GPUs with tensor parallelism (TP=32).