Meta Muse Spark Open is Now Available on LunaRoute

We've added Meta Muse Spark Open to LunaRoute, and from our own usage so far, we're quite impressed. It's a very capable US-developed open-weight model, particularly good at tool use, coding, and long-running agentic tasks.

You can now run it on LunaRoute with Zero Data Retention, on US-based servers, at a fixed monthly cost instead of paying for every token. We also offer a 7-day free trial, so you can try Muse Spark Open alongside the other models we serve and see how they work for your workloads.

Why this model matters

Some of the best open-weight models today come from Chinese labs. DeepSeek, GLM, Qwen, and Kimi have released some impressive models, and we run several of them on LunaRoute.

However, not every organization can use Chinese-developed models. This might be because of compliance requirements, internal security policies, procurement restrictions, or simply a preference to use models developed elsewhere.

It's worth pointing out that using a Chinese-developed model doesn't mean sending your data to China. We host these models ourselves on US servers, without sending requests back to their developers. Still, for some organizations, the origin of the model itself matters, and until recently, finding alternatives with comparable capabilities wasn't always easy.

Muse Spark Open gives these organizations another option, and a very good one.

How good is it?

Muse Spark Open is an 816B parameter Mixture-of-Experts model with around 42B active parameters per token and support for up to 1M tokens of context. Despite its size, it can be surprisingly fast, thanks to its MoE architecture, NVFP4 quantization, and inference optimizations.

Meta's published benchmarks show particularly strong performance in tool use and agentic tasks, including MCP-Atlas and Terminal-Bench. It's competitive with some of the leading open-weight models, although it doesn't win on every benchmark.

Of course, we're living in the age of benchmaxxing. Labs know which benchmarks people care about and optimize heavily for them. That's not necessarily bad, but it also means that ranking first on a benchmark doesn't guarantee a model will perform best on your particular workload.

We've written before about how the model is not the product. The inference engine, quantization, hardware, and agent harness can all make a difference, which is why we care a lot about actually using these models rather than relying entirely on benchmark results.

And from our real-world usage, Muse Spark Open is really good.

We've been particularly impressed with how it handles longer-running tasks, uses tools, and works through problems that require multiple steps. It's a great fit for background agents that need to investigate something, work through a codebase, process documents, or try different approaches before getting to a useful result.

Not every AI task needs someone sitting in front of a chat window waiting for the next response. There's a lot of useful work that agents can do in the background, sometimes for minutes or hours, and that's exactly the kind of workload we've been building LunaRoute for.

Private inference without the token meter

On LunaRoute, Muse Spark Open runs on our US-based infrastructure, with Zero Data Retention by default. We operate the servers ourselves, and your prompts and responses aren't persistently stored or used for training.

Rather than charging for every input and output token, we provide inference capacity at a fixed monthly cost. This is particularly useful for long-running agents, where you want to give them the freedom to explore, retry, make mistakes, and keep working without having to worry about the cost of every additional attempt.

We've also spent a lot of time building a scheduling system that makes better use of GPU capacity. Not every request needs to run immediately, and by intelligently scheduling work that can wait, we can get more useful work out of the same hardware. This is one of the things that allows us to offer fixed-cost inference in the first place.

Shared or dedicated, depending on what you need

Muse Spark Open is now available on our shared infrastructure, alongside the other models we support.

For organizations that need dedicated capacity, we can also deploy Muse Spark Open and other models from our catalog on dedicated infrastructure, ranging from partial GPU nodes to full dedicated servers. These deployments also benefit from our optimized inference stack and scheduling, so you get more than just access to GPUs.

Whether you need stronger isolation, predictable costs, or enough capacity to run a large number of agents, we can work with you on a configuration that makes sense for your needs.

If you're interested in dedicated inference, come talk to us at hello@lunaroute.com.

Otherwise, start your 7-day free trial on LunaRoute and give Muse Spark Open a try.