Muse Spark Open: Free Access, API Pricing and Mac Requirements
Muse Spark Open is Meta's newly released open-weight model for coding and agent tasks. You can download its weights, but running it needs substantial hardware, and OpenRouter lists paid API access. For a first test without that hardware, a hosted provider is the practical route. aiwith has not integrated Muse Spark Open. If you want to work in aiwith today, Kimi K3 and DeepSeek are available alternatives, not access to Meta's model.
Last checked: October 16, 2026. This is a source-checked access guide, not a hands-on benchmark.
What is Muse Spark Open?
The PyTorch announcement by Meta's team dates the release to October 15. It describes a model with 816 billion total parameters, 42 billion active parameters and a 1M-token context window, sharing its base model with Muse Spark 1.3. The open release received additional safety post-training; do not assume every hosted Spark version behaves identically.
The official model card lists text, image, video and audio input, with text output. Meta targets coding, tool use and multi-step agent workflows. Those are intended capabilities, not evidence that it will outperform your current model on your own task.
Is Muse Spark Open free?
There are three different questions behind “free”:
- Can I get the weights? Meta provides BF16 and NVFP4 checkpoints in its Hugging Face collection. They are released under the Muse Model License. Read the license linked from the model card for your intended use; “open weights” is not a claim of unrestricted use.
- Can I run it without paying a provider? Self-hosting avoids a hosted inference bill, but you still supply the hardware, storage and electricity.
- Is hosted chat or API use free? The OpenRouter listing we checked is paid. We have not verified a general free hosted offer for this exact model, including a free official Meta chat endpoint. Account credits or promotions should be checked separately; a free tier for another Muse model does not establish free Spark Open access.
What does OpenRouter charge, and how do I try it?
The OpenRouter model page lists model ID meta/muse-spark-open at $1.25 per million input tokens and $4.25 per million output tokens. We also checked these rates against its public model catalog. Prices and provider availability can change.
For a first test:
- Open that exact model page and check the current price and available providers.
- Use OpenRouter's chat interface if available for your account, or its API with that model ID and an API key. Follow its account and credit requirements.
- Send one small, non-sensitive task before committing a large codebase or long document.
For example, 10,000 uncached input tokens plus 2,000 billed output tokens cost about $0.021 at those rates, before any account-level fees. That is an arithmetic example, not a promise about how many tokens a task will use.
Meta's developer announcement also lists hosted providers including Baseten, Fireworks AI, Groq, Modal, Nebius and Together AI. Their own access rules and billing apply.
Can I run it on a Mac Studio or a normal laptop?
The Mac headline refers to a substantial multi-machine setup. Meta's hardware documentation lists M5 Ultra Mac Studios with either 2 × 512 GB or 4 × 256 GB of unified memory, using MLX and the NVFP4 checkpoint. This is not a documented single-laptop configuration.
The same documentation lists NVIDIA configurations using four Blackwell GPUs for NVFP4 or eight for BF16. The 42B active-parameter figure describes computation per token; it does not mean you only need to store a 42B model.
If you already operate suitable hardware, start with Meta's download guide, then follow the matching runtime recipe in the hardware documentation. Pick the checkpoint before downloading: the guide lists roughly 537 GB for NVFP4 weights and 1.65 TB for BF16, excluding runtime overhead and the KV cache. A download alone does not set up a working inference server.
Is Muse Spark Open available on aiwith?
No, as of this check. An aiwith subscription does not unlock Muse Spark Open. We checked the configured model list and the alternative landing pages below.
If you mainly want help with a document or a coding question in a browser, you can use:
- Kimi K3 on aiwith for a long-document or reasoning task.
- DeepSeek on aiwith for an everyday writing or coding task. Check the model selector for the exact version; the landing-page headline uses broader DeepSeek branding.
Open the model page, check your selected version and allowance, then submit your task. These are task alternatives, not a benchmark ranking or a claim that they match Muse Spark Open's capabilities. If your goal is specifically to evaluate Meta's new release, use a provider that explicitly lists muse-spark-open.