Muse Spark Open: Meta's Open-Weights Model Explained
Muse Spark Open is Meta's open-weights model, released October 15, 2026: 816B parameters, 1M context. See the license, hardware needs and how it ranks.
Muse Spark Open is Meta's open-weights language model, released October 15, 2026, according to its model card. It has 816 billion total parameters, 42 billion active per token, and a 1M-token context window. As of October 17, 2026, Artificial Analysis scores it 39 on its Intelligence Index, which places it below several other open models.
This guide explains what the release is, what its license allows, what hardware it needs, and where it ranks. The model card and license are the primary sources. The leaderboard figures come from Artificial Analysis (AA), read the same day.
Muse Spark Open specs
The table below comes from the model card on Hugging Face.
| Specification | Muse Spark Open | |---|---| | Release date | October 15, 2026 (per the model card) | | Total parameters | 816B (Meta's naming) | | Active parameters | 42B per token | | Architecture | Mixture-of-Experts Transformer with a Perception Encoder | | Language model layers | 62 | | Experts | 256, with 8 active per token | | Perception Encoder | About 1.8B parameters | | Context window | 1M tokens | | Inputs | Text, image, video and audio | | Outputs | Text only | | Weight formats | BF16 and NVFP4 | | Reasoning effort | low, medium, high, xhigh (default xhigh) | | Recommended sampling | Temperature 1.0, top_p 1.0, top_k disabled |
Hugging Face metadata counts about 822 billion parameters in the safetensors files. The gap is mostly the encoder, which is why this guide uses Meta's 816B figure. Source: Muse Spark Open model card.
No Meta blog post announcing the release was available to read for this guide, so the release date comes from the model card alone.
Is Muse Spark Open open source?
Meta releases the weights, but the model is not under a standard open-source license. It is under the Muse Model License, a custom license with conditions. This guide calls it "open weights" for that reason. The summary below is a plain-language reading of the license text, not legal advice.
- Allowed: use, copy, modify (including fine-tuning), distill, distribute and sell.
- Required: include the license and the Usage Policy with copies and derivatives, and display "Built with Muse Spark Open" in your product UI, documentation, website or settings.
- Outputs: Meta says it claims no ownership of the outputs.
- Revenue clause: if the licensee and its affiliates together earn more than $20 million USD in any 12 consecutive months, and they offer the model or a derivative as a Model-as-a-Service product (third-party access through an API or similar), they must notify Meta and sign a separate written agreement before doing so.
- No user threshold: the clause has no monthly active user test.
- Exempt: internal use, research, other commercial activity that is not Model-as-a-Service, and using the model as a component of your own product.
- Usage Policy: prohibits illegal activity, harassment, unlicensed professional practice, unauthorized inference of sensitive personal data, malware, weapons, misuse of critical infrastructure, fraud and disinformation, among other items. The model is not intended for users under 18.
- Access: the Hugging Face page uses manual approval.
The license text read for this guide has no explicit geographic restrictions.
Source: Muse Model License.
How much hardware does Muse Spark Open need?
The model card sets the bar high. Running it at NVFP4 needs at least 700GB of VRAM. The BF16 version needs about 2TB. Meta also says the model is optimized for two Mac Studios with M5 Ultra chips, using MLX. The card lists these inference frameworks: SGLang, vLLM, Hugging Face Transformers, TensorRT-LLM, Unsloth and Llama.cpp. Third-party inference providers also offer the model.
In practice, this is not a model for a single consumer GPU or a laptop. Most teams will run it through a hosting provider or a large multi-GPU server.
How Muse Spark Open ranks on open-weights models
Artificial Analysis lists the open-weights models below on its Intelligence Index, as of October 17, 2026. Parameters are total and active.
| Model (effort) | Intelligence Index | Parameters (total / active) | |---|---|---| | MiMo-V2.6-Pro | 46 | 1.0T / 42B | | GLM-5.3 (Max) | 45 | 753B / 40B | | Step 5 Preview | 44 | 600B / 27B | | Kimi K3 (Max) | 44 | 2.8T / 104B | | GLM-5.3-Flash | 42 | 320B / 18B | | DeepSeek V4.1 Flash (Max) | 39 | 552B / 16B | | Muse Spark Open (Xhigh) | 39 | 816B / 42B | | Qwen3.8 27B (Xhigh) | 34 | not stated in this table |
Source: Artificial Analysis open-weights models. Meta's card lists Kimi K3 at 45, while this table uses AA's 44.
Muse Spark Open is mid-table among open models on this index. It is not the top open-weights model. Meta's proprietary Muse Wonder, released October 12, 2026, is a separate closed model. AA scores it 52, and its weights are not public. The top open-weights model on AA's index is 12 points below the top closed model, Claude Opus 5.5 at 58. That gap is derived from the two AA scores.
Meta's own benchmark claims
Meta's model card reports results for Muse Spark Open at xhigh effort and compares it with other open models. These are vendor-reported numbers, not independent tests. The subset below is taken from that card.
| Benchmark (Meta-reported) | Muse Spark Open | Kimi K3 | GLM-5.3 | |---|---|---|---| | Terminal-Bench 2.1 (Terminus 2) | 76.4 | 85.0 | 83.9 | | GPQA Diamond | 89.0 | 93.5 | not listed | | OSWorld Verified | 80.4 | 84.8 | not listed |
The card also reports Muse Spark Open at 89.2 on MCP-Atlas and 50.0 on ZeroBench with tools. The comparison figures read for this guide do not include those rows, so this guide does not compare them. On the rows above, Muse Spark Open trails both rivals.
Muse Spark Open vs Kimi K3, GLM-5.3, DeepSeek V4.1 Flash and Qwen3.8
License terms matter as much as scores for most teams. The table lists each model's license as stated on its Hugging Face page.
| Model | Size (total / active) | License | Context | |---|---|---|---| | Muse Spark Open | 816B / 42B | Muse Model License (custom) | 1M tokens | | Kimi K3 | 2.8T / about 104B | Kimi K3 License (custom) | 1M tokens | | DeepSeek V4.1 Flash | 552B / 16B (decoding) | MIT | 1M tokens | | MiMo-V2.6-Pro | 1.0T / 42B | MIT tag on Hugging Face | 1M tokens | | Qwen3.8-2.4T-A95B | 2.4T / 95B | "qwen3.8-max" (license text not read) | 262,144 native, extendable to 1,010,000 | | GLM-5.3 | 753B / 40B | "glm-5.3" (license text not read) | not read for this guide |
The Kimi K3 license is not a simple permissive license. A company that offers Kimi K3 as a paid third-party service with more than $20 million in revenue over 12 months must sign an agreement with Moonshot AI first. Products with more than 100 million monthly active users or $20 million in monthly revenue must display "Kimi K3." Use through Moonshot AI's own products or certified partners is exempt from the service clause. Source: Kimi K3 license.
Only the MIT-tagged models in this table, DeepSeek V4.1 Flash and MiMo-V2.6-Pro, were confirmed as MIT in the pages read. Do not read "open weights" as "free for commercial use" without reading each license.
Safety notes from Meta
Meta reports extra safety post-training for the open version. It says the model refused 99.2% of jailbreak attempts on an internal cybersecurity benchmark, and 94.8% across combined public benchmarks, with no system safeguards. It also reports over 98% on BioTIER. These are Meta's results for the tested checkpoint, and they have not been independently verified.
The card tells deployers to add system-level guardrails and not to expose raw chain of thought. It also lists limitations:
- Outputs can be inaccurate or biased.
- Multi-step reasoning can fail in new scenarios.
- Quantized versions may behave differently from the full model.
- Very short or low-resolution video may be rejected. Meta recommends at least 448 pixels on the short side.
Who should consider Muse Spark Open
It fits teams that have large GPU capacity or a hosting partner, need a model that accepts image, video and audio input, and want to fine-tune or distill under a custom license. It is a weaker fit for teams that need speech output, need a small model, or expect to offer the model as a hosted service above the $20 million revenue threshold.
Where voice fits
Muse Spark Open returns text only. A spoken version of its answers needs a separate text to speech model. OndaLabs is an AI voice studio. Its text to speech offers 30 prebuilt voices, usable in every supported language, and 37 language pages, with Onda Lite at 1 credit per character on all plans and Onda Flash at 1.5 credits per character on paid plans. See the language pages and pricing for details.
For the wider field, read The best AI models right now: October 2026 roundup. For a closed frontier comparison, see Claude Opus 5.5: pricing, benchmarks and what's new.
Frequently asked questions
What is Muse Spark Open?
Muse Spark Open is Meta's open-weights language model, released October 15, 2026, according to its model card. It has 816B total parameters, 42B active per token, and a 1M-token context window. It accepts text, image, video and audio input and returns text.
Is Muse Spark Open really open source?
It is open weights under a custom license, the Muse Model License, not a standard open-source license. The license requires attribution, a usage policy and a separate agreement once a company's Model-as-a-Service revenue passes $20 million in 12 months. Read the license before you build a product on it.
Can I use Muse Spark Open commercially?
The license allows commercial use, including selling products, but it has conditions. You must include the license and usage policy, display "Built with Muse Spark Open," and sign a separate agreement if you offer the model as a hosted service above the revenue threshold. Internal use and products that embed the model are exempt from that threshold. This is not legal advice.
How much hardware do you need to run Muse Spark Open?
Meta's card gives 700GB of VRAM as the minimum for the NVFP4 version, and about 2TB for BF16. It is also optimized for two Mac Studios with M5 Ultra chips using MLX. For most teams, a hosting provider is the practical route.
How does Muse Spark Open compare with Kimi K3 and GLM-5.3?
On the Artificial Analysis index on October 17, 2026, Muse Spark Open scored 39, behind MiMo-V2.6-Pro (46), GLM-5.3 (45), and Kimi K3 (44). On Meta's card, it trails Kimi K3 and GLM-5.3 on Terminal-Bench 2.1, and trails Kimi K3 on GPQA Diamond and OSWorld Verified. Each model has different license terms.
What is the best open-weights AI model right now?
On the Artificial Analysis index on October 17, 2026, MiMo-V2.6-Pro scored highest among open-weights models, at 46. Scores move as new models arrive, so check the leaderboard before you decide. License terms and hardware needs can matter more than a single score.