Test Your MCP Server With Meta Muse
Free2 Meta Muse models
Paste a server URL and watch Muse Spark 1.2 or the open Muse Glimmer 30B call your tools for real — arguments, results, errors and all. No API key, no local install, no Meta account.
Why test with Meta Muse?
The Muse line is the strongest showing on MCP-Atlas, Scale AI’s Model Context Protocol benchmark: Muse Spark 1.1 ranks first at 88.1%, ahead of Claude Opus 5 at 85.8%. Spark 1.2 adds a 1M-token context and parallel function calling for orchestration that will not fit on a laptop. Muse Glimmer 30B is the Apache 2.0 model distilled from it — Meta reports 75.5 on the same benchmark — and it fits under 20GB at 4-bit, so a good result here means something unusual: you could self-host the model driving your tools.
How Meta Muse reaches MCP tools
- An MCP client that converts your tool definitions into OpenAI-shaped function calls — what MCP Playground does here, and the path with no setup
- A local runtime for Glimmer specifically (Ollama, LM Studio, llama.cpp, MLX, vLLM, SGLang) driving a locally-running MCP server, so no tool call ever leaves your machine
- Muse Spark’s API for multi-agent setups, where it runs as the planning agent or as a parallel subagent
Meta has not published a tool-calling wire format for either Muse model. The model card and developer pages assert reliable tool use and precise schema invocation, but no serialization is specified, and the reference runtimes simply expose an OpenAI-compatible surface. So the honest statement is this: routed through here, both models are OpenAI-shaped function callers like every non-Anthropic model in the catalogue, and any claim you read about how Muse specifically handles nested arguments is currently untested rather than documented. That makes your own server the only authority — which is the point of running it.
What goes wrong with Meta Muse specifically
- Reasoning effort changes the model you are testing. Glimmer exposes low, medium, high and xhigh, and Meta recommends high or xhigh for agentic work — a disappointing run at default effort is not the model the benchmarks describe.
- Sampling defaults are unusually high: temperature 1.0, top_p 0.95, top_k 64. Carrying over a temperature 0.2 habit from another vendor is a real source of erratic tool selection.
- Glimmer leads on agentic benchmarks but loses SWE-Bench Verified to Qwen3.6-27B (76.0 vs 77.2), on Meta’s own reported numbers. It is tuned for choosing and sequencing tools, not for single-shot code editing — judge it on the former.
- Spark and Glimmer are not interchangeable on modality: Spark accepts audio, Glimmer does not, and Glimmer treats video as individual frames. A pipeline that assumes parity between teacher and student will break.
- Glimmer is tuned for calling tools through a protocol, not for driving a machine. Qwen3.6-27B beats it clearly on TerminalBench 2.1 (60.7 to 51.7) and OSWorld-Verified (75.6 to 65.9) on Meta’s own numbers. Strong MCP results do not transfer to terminal or desktop control.
- On Siren AgentDojo, Meta reports a 28.4% prompt-injection attack success rate for Glimmer (lower is better; Gemma4-31B is 25.6%, Qwen3.6-27B 40.3%). If your tools return content you did not author, roughly one in four injection attempts lands — pick the model on the margin, but keep the controls on your side.
Meta claims Glimmer is trained so that when a tool call fails or returns something unexpected, it diagnoses the error and retries rather than halting. Point it at the error mock server and check: a model that lives up to that reads the error, adjusts its arguments and calls again, while most models apologise and stop. It is the most falsifiable claim in the launch materials and it takes one run to settle. All the mock servers →
Meta Muse models available here
| Model | Model ID | Credits / run |
|---|
| Muse Spark 1.2 | meta/muse-spark-1.2 | 4 |
| Muse Glimmer 30B | meta/muse-glimmer-30b | 2 |
Frequently asked questions
Can Meta Muse models use MCP servers?
Yes. Neither Muse model implements MCP at the API level — only Anthropic does — but any MCP client can translate your server’s tools into function-calling format and drive them. That is what happens when you paste a URL here: your server stays unchanged and the client does the translation. Muse Spark’s first-place ranking on MCP-Atlas suggests the line handles that translation unusually well.
What is MCP-Atlas, and why does it matter?
It is Scale AI’s benchmark for Model Context Protocol tool use — 1,000 tasks across 36 real MCP servers and 220 tools, at three to six tool calls each, where the model must find the right tool, call it correctly, handle errors and combine results. Muse Spark 1.1 ranks first on the public leaderboard at 88.1%. Meta separately reports 75.5 for Glimmer against 62.5 for Qwen3.6-27B and 54.2 for Gemma4-31B; that figure is a vendor comparison, not a leaderboard entry.
Do I need a Meta account or API key to test this?
No. Runs here go through MCP Playground’s own routing, so you can compare Muse against other models without holding an account with any vendor. You only need your own key if you are wiring Muse into your product afterwards.
Should I use Muse Spark 1.2 or Muse Glimmer 30B?
They solve different problems. Glimmer is Apache 2.0 and runs locally at roughly 20GB when quantized, so pick it when tool calls must stay on your hardware or when you are checking whether an open 30B is good enough to replace a frontier model on your tools. Pick Spark 1.2 when the task genuinely needs its 1M-token context or parallel function calling — whole-repository work, long debugging sessions, multi-agent orchestration.