AIREITERMuse Spark Open requires at least two top-memory Mac Studios in published demonstrations, so most developers should test it through a hosted API first.
Is Muse Spark Open worth using?
Muse Spark Open is worth evaluating for private, long-running agent workflows, especially when a team needs downloadable weights and broad runtime support. It is not the obvious default for ordinary coding tasks because its infrastructure burden is extreme and Meta’s reported coding scores trail several open-weight alternatives.
| Situation | Recommendation | Reason |
|---|---|---|
| Testing agent or tool-use quality | Use a hosted provider | No cluster purchase or distributed MLX setup |
| Private on-premises workflows | Consider self-hosting | Weights stay under organizational control |
| Routine code generation | Compare smaller models first | Muse Spark Open does not lead the cited coding tests |
| One consumer GPU or one Mac | Do not self-host | The 4-bit checkpoint leaves no room for inference workspace |
| Commercial model API business | Review the license first | A revenue threshold and separate-agreement clause may apply |
| Full 1M-token workload | Validate with representative data | Published Apple tests stop at 65,536 tokens |
The launch is official rather than a preview. Meta released Muse Spark Open on October 15, 2026, with downloadable weights and access through seven named inference providers.
What Meta actually released
Muse Spark Open is an open-weight, natively multimodal mixture-of-experts model built for coding, software engineering, tool use, and long-horizon agents. It accepts text, image, audio, and video inputs and produces text output.
| Specification | Published value |
|---|---|
| Total parameters | 816B |
| Active parameters per token | 42B |
| Context window | 1M tokens |
| Weight formats | BF16 and NVFP4 |
| Training-language coverage | More than 100 languages |
| Hosted API providers named by Meta | 7 |
| Supported runtimes named by Meta | 7 |
The 42B-active design means the model does not compute across all 816B parameters for each token. Storage remains the harder constraint: Apple reports that native BF16 parameters require more than 1.6TB of memory, while the 4-bit NVFP4 checkpoint occupies about 493GB before inference workspace and context memory.
“Open-weight” is the accurate term. Developers can download, modify, fine-tune, distill, and distribute the model under the Muse Model License, but the license includes attribution, usage restrictions, and special terms for some model-as-a-service businesses. Those conditions do not match a conventional permissive open-source license.
The local deployment reality
Local Muse Spark Open deployment is technically credible but financially and operationally demanding. Meta demonstrates two M5 Ultra Mac Studios with up to 512GB of unified memory each; Apple’s distributed MLX report measures two- and four-machine configurations.
Why one 512GB Mac is not enough
One 512GB Mac Studio cannot hold the approximately 493GB NVFP4 checkpoint plus the memory needed for intermediate inference work. Apple therefore distributes the model across multiple machines using MLX and RDMA over Thunderbolt 5.
Apple’s test used macOS 26.2, MLX 0.30.1 or later, the jaccl communication backend, and two Thunderbolt cables between peers. It measured 8,192- and 65,536-token contexts and generated 128 tokens at batch size one.
| Apple configuration, 8,192-token prompt | Time to first token | Decode speed |
|---|---|---|
| 2 M5 Ultra, pipeline parallel | 8.4 s | 15.74 tok/s |
| 2 M5 Ultra, tensor parallel | 8.2 s | 24.99 tok/s |
| 4 M5 Ultra, tensor parallel | 4.8 s | 36.73 tok/s |
The four-machine figure is the performance headline, not the minimum deployment. Meta’s two-machine setup shows that the model fits and runs; Apple’s measurements show how much topology changes the experience.
Pipeline versus tensor parallelism
Pipeline parallelism splits the model by layer depth and is the simpler way to fit the weights across two machines. It adds relatively little communication, but sequential layer execution limits decode acceleration.
Tensor parallelism splits each layer across machines. Apple measured 24.99 tokens per second on two M5 Ultra systems, 59% faster than the 15.74-token pipeline result under the 8,192-token test. The trade-off is heavier communication; at 65,536 tokens, Apple found that two-machine tensor prefill could be slower than pipeline prefill.
Performance: strong on two agent tests, weaker on coding
Muse Spark Open leads two published agent and tool-use comparisons, not every agentic benchmark. The values below are Meta-reported rather than independently reproduced, so they support shortlisting rather than a production decision.
| Benchmark | Muse Spark Open | Strongest cited comparison |
|---|---|---|
| JobBench | 61.8 | GLM 5.3: 61.4 |
| MCP-Atlas | 89.2 | Qwen3.8: 84.5 |
| Terminal-Bench 2.1 | 76.4 | Kimi K3: 85.0 |
| DeepSWE 1.1 | 60 | Kimi K3: 68 |
| SWE-Atlas-QnA | 58 | Kimi K3: 66 |
| GPQA Diamond | 89.0 | Kimi K3 and Qwen3.8: 93.5 |
Muse Spark Open leads JobBench by 0.4 points and MCP-Atlas by 4.7 points over the next values shown. On Terminal-Bench 2.1, it trails Kimi K3 by 8.6 points and GLM 5.3 by 7.5 points.
Meta also describes two local demonstrations: changing and building the open-source Ice Cubes iOS app, and operating DaVinci Resolve to inspect footage and produce vertical crops. Neither demonstration reports completion rate, elapsed time, intervention count, or reproducible task files, so they show capability rather than reliability.
Independent production evidence for the released Open checkpoint remains limited.
License terms that can block deployment
The Muse Model License permits broad use but imposes conditions that product and legal teams should examine before deployment. The practical blockers are attribution, the model-as-a-service threshold, and prohibited-use rules.
| License issue | Practical effect |
|---|---|
| Attribution | Products and services must prominently show “Built with Muse Spark Open” |
| Model-as-a-service threshold | Businesses above $20M aggregate gross revenue in a consecutive 12-month period must notify Meta and obtain a separate agreement before offering covered model access |
| Internal or embedded product use | The cited model-as-a-service provision does not apply in the same way to internal use or a model powering a non-model-service product |
| Usage restrictions | The policy restricts areas including malware, weapons, critical infrastructure, fraud, and unauthorized impersonation |
| Age statement | The model is not intended for people under 18 |
The $20 million clause is not an API fee. It is a licensing threshold tied to qualifying model-as-a-service activity. Embedding Muse Spark Open inside an application may require a different analysis from reselling it through an inference endpoint.
How to access Muse Spark Open now
Muse Spark Open is available through hosted APIs and downloadable model catalogs. Meta names Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI for hosted access; weights are listed through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.
Supported runtimes include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp, and Unsloth. Meta reports a 3% end-to-end improvement from its cited Helion-kernel test on AMD MI350 hardware.
There is no single official hosted price because providers set their own routes and terms. Compare token prices, precision, context limits, reasoning settings, regional routing, data-retention and model-training policies, and whether the provider serves the released Open checkpoint.
Muse Spark Open FAQ
Is Muse Spark Open actually open source?
No. The weights are downloadable and modifiable, but the license contains attribution, commercial, and usage restrictions. “Open-weight” is the safer description.
Can Muse Spark Open run on one Mac Studio or GPU?
Not in Apple’s published setup. The NVFP4 checkpoint is about 493GB before workspace, while BF16 parameters require more than 1.6TB; practical deployment therefore needs multiple machines or accelerators.
How much does the Muse Spark Open API cost?
Pricing depends on the hosted provider. Meta names seven API partners but does not publish one universal token price.
Is Muse Spark Open better than Kimi K3 or GLM 5.3?
Muse Spark Open leads JobBench and MCP-Atlas in the cited results, while Kimi K3 and GLM 5.3 lead several coding-oriented tests. Evaluate the same workload rather than relying on an overall label.
What to test before self-hosting
Before buying hardware, run the same representative agent tasks through two hosted providers. Record task completion, human interventions, time to first token, total latency, output cost, data-retention terms, and checkpoint identity; self-host only if privacy or control outweighs the measured operational cost.