Muse Spark Open
- Provider
- Meta
- Status
- available
- Context
- 1,000,000 tok
- Price
- $1.25 / $4.25 /MTok
Muse Spark Open is an open-weight model from Meta, released on 15 October 2026: a mixture-of-experts model with 816 billion parameters, 42 billion active, and a 1M-token context (Meta). It is the open-weight release Meta promised when it made Muse Spark proprietary, and Meta calls it the most capable open-weight model yet from a US company. The weights are free to download under the Muse Model License, which Meta says has “limited commercial terms”.
Artificial Analysis scored it 39 on its Intelligence Index v4.3 at launch, and 25 on v5.0 at xhigh reasoning when it launched that version on 27 October, then the highest-scoring open-weight model developed in the US. Meta published no SWE-bench figure.
Quick specs
| Muse Spark Open | |
|---|---|
| Released | 15 October 2026 |
| Architecture | Mixture-of-experts; 816 billion parameters, 42 billion active |
| Modalities | Text, images, speech and video in, text out (Artificial Analysis) |
| Context window | 1M tokens |
| Hosted price | $1.25 in / $4.25 out per MTok, median across hosts (Artificial Analysis) |
| AA Intelligence Index | 39 at launch (v4.3); 25 on v5.0 (from 27 October, xhigh) |
| Cost per AA index task | $2.02 |
| Output speed | 139 tokens a second (Artificial Analysis) |
| Licence | Muse Model License |
What Muse Spark Open is
Meta took its Muse Spark line proprietary in April 2026, and Mark Zuckerberg said open-weight releases would follow. Muse Spark Open is that release. Meta describes it as “a natively multimodal model”; Artificial Analysis lists text, image, speech and video input with text output. It is a separate model from the Muse Spark versions on the Meta Model API, which stay closed.
It is Meta’s second open-weight model since the pivot, after Muse Glimmer, a 30-billion-parameter model released under Apache 2.0 in August.
Benchmark performance
| Measure | Muse Spark Open | Context |
|---|---|---|
| Intelligence Index v4.3 | 39 | Scored at launch, before v5.0 |
| Intelligence Index v5.0 | 25 (xhigh) | In late October, on the same version: Mistral Large 4 26; DeepSeek V4.1-Flash and GLM-5.3 32 |
| Cost per index task | $2.02 | |
| Output tokens used across the index | 91M | |
| Output speed | 139 tokens a second | Time to first token 0.73 seconds |
Source: Artificial Analysis, independent. Artificial Analysis rebuilt its index as v5.0 on 27 October, and scores on it run lower than on v4.3, so the two readings are not comparable.
Meta’s launch post links an evaluation methodology report but prints no benchmark figures, and Meta has published no SWE-bench Pro or SWE-bench Verified score.
Where to run it
The weights are on Hugging Face and through Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. API access is through third-party hosts: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI (Meta). Meta does not list it on its own Meta Model API.
Meta publishes no price. Artificial Analysis puts the median host price at $1.25 per million input tokens and $4.25 per million output tokens.
The licence
Meta releases the weights under the Muse Model License “with limited commercial terms, which allow many developers to use the model for free”. Meta’s post gives no thresholds. DeepLearning.AI reports that companies with more than $20 million in revenue that sell access to the model need a separate agreement with Meta. Muse Glimmer is under Apache 2.0 and Llama 4 under the Llama Community Licence, so Meta’s open models now carry three different licences.
How it compares
vs Mistral Large 4
When Mistral released its weights on 29 October, Mistral Large 4 scored 26 on the v5.0 index, a point above Muse Spark Open. It has 1.05 trillion parameters (52 billion active) and a 1M-token context. Mistral’s own API charges $0.70/$2.10 per million tokens, and its modified MIT licence asks only businesses selling the model as a hosted service for a commercial licence.
vs DeepSeek V4.1-Flash
On v5.0 in late October, DeepSeek V4.1-Flash scored 32, seven points higher. It has 552 billion parameters under the MIT licence and costs $0.15/$0.60 off-peak on DeepSeek’s API.
vs Muse Glimmer
Muse Glimmer is Meta’s small open model: 30 billion dense parameters, a 131K context, Apache 2.0, and built to run on a single 24GB consumer GPU. Muse Spark Open is far larger and needs data-centre hardware or a host.
Known limitations
No published benchmarks from Meta. Apart from the Artificial Analysis index, there are no figures to judge it by, and no SWE-bench score.
Licence with conditions. The Muse Model License is not an open-source licence, and Meta has not published its thresholds in the launch post.
Size. At 816 billion parameters it is a multi-GPU self-host.
Frequently asked questions
When was Muse Spark Open released?
Meta released Muse Spark Open on 15 October 2026, with the weights on Hugging Face and API access through third-party hosts the same day.
Is Muse Spark Open open source?
It is open-weight, not open source. The weights are downloadable under the Muse Model License, which Meta says has limited commercial terms. DeepLearning.AI reports that companies with more than $20 million in revenue that sell access to the model need a separate agreement with Meta.
How much does Muse Spark Open cost?
The weights are free to download. Meta sets no API price; Artificial Analysis lists a median of $1.25 per million input tokens and $4.25 per million output tokens across the hosts serving it, and $2.02 per completed Intelligence Index task.
How many parameters does Muse Spark Open have?
816 billion in total, with 42 billion active per token, according to Meta.
How does Muse Spark Open score on independent tests?
Artificial Analysis scored it 39 on its Intelligence Index v4.3 at launch. On v5.0, which it launched on 27 October, it scored 25 at xhigh reasoning, then the highest-scoring US-developed open-weight model, a point below Mistral Large 4 when Mistral released its weights on 29 October.
Is Muse Spark Open the same as Muse Spark?
No. The Muse Spark models on the Meta Model API are proprietary. Muse Spark Open is a separate open-weight model in the same family.