Meta just released the weights for Muse Spark Open, and this one is big. The company calls it the most capable US open-weight model to date. It landed on October 15, 2026, only three days after Meta launched its closed Muse Wonder models.
So what do you actually get? In short, you get a huge multimodal model that you can download, tune and run yourself. However, the license, the hardware bill and the benchmarks all deserve a closer look. Let me walk through each one.
What Muse Spark Open Actually Is
According to Meta’s announcement, the model has 816 billion total parameters. Yet only 42 billion are active for each token, because it uses a mixture-of-experts design. It also supports a context window of 1 million tokens.
Meta built it for coding and agentic work. In particular, it targets long-horizon tasks, complex software engineering and multi-step reasoning. The model is natively multimodal, too. It accepts text, images, audio and video, but it only outputs text.
The PyTorch team’s post adds useful context. This release shares the same base model as Muse Spark 1.3. On top of that, it received extra safety post-training. The goal is to make it more robust against misuse when it runs without system guardrails.
In other words, this is not a stripped-down toy. Instead, it is an open version of Meta’s own production model, with extra safety work for life outside Meta’s servers.
Running It on Two Mac Studios
The headline demo is about your desk, not a data centre. Meta says this is the first model in its class to run on a desktop setup of two Apple M5 Ultra Mac Studios. The latest MLX framework makes that possible.
Here is the logic. Each M5 Ultra Mac Studio can give its GPU up to 512 GB of unified memory. Therefore, two machines can hold the entire model. Meta says the pair behaves like one computer with a terabyte of memory.
Meta also showed agent demos that ran fully on-device. For example, Muse Code powered by the model shipped real changes to Ice Cubes, an open-source Mastodon client. In another demo, it drove DaVinci Resolve and reviewed footage frame by frame to find the best moments.
Still, there is a catch. VentureBeat notes that Apple’s 512 GB configuration does not arrive until late October. Its price is not announced yet. Moreover, the model card lists at least 700 GB of accelerator memory for the NVFP4 version. So “desktop” here still means a serious investment.
Muse Spark Open Benchmarks: Strong in Spots
Meta’s own table shows clear wins. For instance, the model scores 61.8 on JobBench and 89.2 on MCP-Atlas. Both results beat the rivals Meta listed, including GLM 5.3, Kimi K3 and Qwen3.8.
However, the picture changes on coding. On Terminal-Bench 2.1, it scores 76.4. By contrast, Kimi K3 reaches 85.0 and GLM 5.3 reaches 83.9. Similarly, it trails both of them on DeepSWE 1.1.
Meta also reports an Artificial Analysis Intelligence Index score of 39. Yet VentureBeat points out that several Chinese open-weight models have higher published scores. As a result, “most capable US open-weight model” is a narrower claim than “best open model in the world.”
My take is simple. The tool-use numbers are the real story. If your agents call many tools through MCP, Muse Spark Open deserves a test. If you mostly need terminal coding, compare it with the rivals first.
The License Fine Print
Meta releases the weights under the Muse Model License. According to Meta, its limited commercial terms let many developers use the model for free. Even so, the details matter.
VentureBeat breaks down the main rule. Some companies earn more than $20 million in gross revenue over 12 months, counting affiliates. Those companies need a separate agreement with Meta before they offer the model as a service. On the other hand, internal use, research and features inside your own product are exempt.
There are other conditions as well. Products must show the line “Built with Muse Spark Open” in a permitted place, such as documentation or settings. In addition, the Hugging Face download form asks for your full legal name and date of birth.
So yes, anyone can apply to download the weights. But this is an open-weight release under a custom license. It is not an Apache-style free-for-all.
Where to Get It and What It Costs
Access is wide from day one. Meta lists API access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Meanwhile, the weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.
Pricing depends on the host. For example, VentureBeat reports Together AI rates of $1.25 per million input tokens and $4.25 per million output tokens. That matches Meta’s own hosted Muse Spark pricing.
Context limits vary, too. According to VentureBeat, GroqCloud lists a limit of 131,072 tokens. Fireworks, however, lists the full 1,048,576 tokens. Therefore, check your provider before you plan a giant-context workflow.
What about tuning? PyTorch’s TorchTitan now supports the model for SFT, LoRA and reinforcement learning. Plus, Meta ships Helion GPU kernels in SGLang. These kernels bring 1.03x to 3.95x kernel-level speedups on NVIDIA and AMD hardware.
Should You Try Muse Spark Open?
It depends on what you build. First, if you need control over data and deployment, you get US-developed weights that you can host yourself. Second, if your agents lean on tool calling, the MCP-Atlas result looks promising.
On the flip side, the hardware bill is real. Rival models still lead on some coding tests. Also, larger companies that sell model access should read the revenue clause with care.
If you read my post on the Muse Wonder launch, the strategy now looks clear. Meta keeps its strongest model closed. At the same time, it opens a capable sibling for builders. Together, the two releases put Meta on both sides of the open-versus-closed debate.
Bottom line: Muse Spark Open is not the top open model on every chart. Nevertheless, it is one of the most interesting US open-weight releases of 2026. For most AI teams, it is worth a weekend test.
No comment