๐ TL;DR
- Muse Spark 1.1 ranks #1 on Scale AI's MCP-Atlas leaderboard at 88.1% โ ahead of Claude Opus 5 (85.8%) and Claude Fable 5 (83.3%).
- Muse Glimmer 30B is Apache 2.0 and runs locally โ under 20GB at 4-bit, so it fits a 24GB machine.
- Glimmer is distilled from Spark via logit distillation. Teacher and student, not two sizes of one model.
- Meta reports 75.5 on MCP-Atlas for Glimmer vs 62.5 (Qwen3.6-27B) and 54.2 (Gemma4-31B) โ a vendor figure, not a leaderboard entry.
- Glimmer loses badly at terminal and desktop control. Qwen wins TerminalBench 2.1 60.7 to 51.7 and OSWorld-Verified 75.6 to 65.9.
- Neither model's tool-calling wire format is documented. Through OpenRouter both are OpenAI-shaped function callers.
Meta shipped two models a week apart in August 2026, and the relationship between them is the story. Muse Spark 1.2 is closed, API-only, with a million-token context. Muse Glimmer 30B is Apache 2.0, open-weights, and distilled from Spark โ it runs on a single consumer GPU.
If you build on MCP, the student is the more interesting release. Here is why, and where the marketing stops matching the data.
MCP-Atlas is a real benchmark, and Muse tops it
Most launches give you MMLU and a coding score. Neither predicts whether a model will drive your MCP server without mangling arguments.
MCP-Atlas is different. It is Scale AI's benchmark, built specifically for Model Context Protocol tool use: 1,000 tasks across 36 real MCP servers and 220 tools, three to six tool calls per task. The model has to discover the right tool from a limited set, call it with correct parameters, handle errors, coordinate across servers, and synthesise an answer.
On Scale's public leaderboard:
| Rank | Model | Pass rate |
|---|---|---|
| 1 | Muse Spark 1.1 | 88.1% |
| 2 | claude-opus-5 (xhigh) | 85.8% |
| 2 | gemini-3.5-flash (high) | 83.6% |
| 2 | Claude Fable 5 | 83.3% |
| 2 | kimi-k3 (max) | 82.3% |
| 2 | claude-opus-4-8 (max) | 82.2% |
Two caveats, both load-bearing. The leaderboard was last updated April 2026, so the ranked entry is Spark 1.1, not 1.2. And Glimmer, which launched in August, is not on it at all. This is evidence the Muse line is strong at MCP work โ not a verified score for the build you are calling today.
What Meta reports for Glimmer
Glimmer's numbers are from Meta's own launch table, so read them as a vendor comparison. Conditions matter: Glimmer at High Reasoning, the rivals in Thinking Mode. All best-effort configurations, not defaults.
| Benchmark | Muse Glimmer-30B | Qwen3.6-27B | Gemma4-31B |
|---|---|---|---|
| MCP Atlas | 75.5 | 62.5 | 54.2 |
| DeepSearch QA | 74.6 | 71.1 | 61.7 |
| ฯยฒ-Banking | 23.5 | 16.7 | 15.1 |
| GAIA2 | 43.3 | 40.0 | 36.4 |
| SWE-Bench Pro | 51.2 | 50.2 | 36.9 |
| SWE-Bench Verified | 76.0 | 77.2 | 66.6 |
| TerminalBench 2.1 | 51.7 | 60.7 | 43.4 |
| OSWorld-Verified | 65.9 | 75.6 | 58.5 |
The pattern is sharper than "Glimmer wins."
Glimmer's advantage is protocol-shaped agentic work โ discovering tools, invoking schemas correctly across a long workflow, recovering from failures. On MCP-Atlas the gap over Qwen is 13 points; on ฯยฒ-Banking it is a rout.
But Qwen3.6-27B takes SWE-Bench Verified, TerminalBench 2.1 and OSWorld-Verified, and those last two are not close. The honest reading: Glimmer is tuned for calling tools through a protocol, not for driving a terminal or a desktop. Strong MCP results do not transfer to computer use.
The number nobody puts in a headline
Meta also publishes safety results. On Siren AgentDojo, a prompt-injection benchmark, Glimmer records a 28.4% attack success rate at 94.2 utility. Gemma4-31B is better at 25.6%; Qwen3.6-27B is worse at 40.3%.
Lower is better, and none of these are comfortable. Roughly one in four injection attempts lands against the best of the three.
If your MCP server returns content that originated outside your control โ search results, file contents, third-party API responses โ that content reaches the model, and these are the empirical odds it can steer it. Model choice moves the margin. It does not remove the problem. The controls that matter stay on your side of the boundary.
Glimmer's architecture
A dense causal transformer โ not MoE โ at ~29.6B parameters including a 1.8B ViT-G/14 perception encoder. 52 layers, hidden dim 6,656, SwiGLU at 19,968 intermediate width.
Two choices stand out if you plan to run it:
- GQA at 16:1 โ 32 query heads against 2 KV heads. Aggressive, and it is what keeps the KV cache small enough for long agentic sessions on consumer memory.
- A repeating [Local, Local, Local, Global] attention pattern with a 2,048-token sliding window. Three cheap local layers per global one.
Context is 131,072 tokens by default, extendable to 262,144. Sampling defaults are temperature 1.0, top_p 0.95, top_k 64 โ carrying over a temperature 0.2 habit from another vendor is a real source of "it did worse for me."
Full precision needs 55GB+. Quantized to ~4-bit it drops under 20GB, leaving headroom for KV cache and the perception encoder in a 24โ32GB envelope. Meta reports "minimal to no degradation on agentic tasks" from that compression.
DFlash speculative decoding ships with it: a lightweight drafter proposes token blocks, the main model verifies in parallel. Output quality is identical by construction; measured speedup is 3.1x on an RTX 5090, 1.8x on an M5 Max, 1.5x on an M4 Max.
Runtimes at launch: Ollama, LM Studio, llama.cpp, MLX, ExecuTorch, vLLM, SGLang.
The one claim worth testing yourself
Meta describes three training phases โ logit distillation from Spark in pre-training, extended-context agent data with reasoning traces in mid-training, then SFT blended with on-policy distillation and RL.
Buried in that is the most falsifiable claim in the launch materials:
"When a tool call fails or returns an unexpected result, the model is trained to diagnose the error and retry rather than halt."
That is the behaviour separating a model you can leave running from one that needs a human every time an endpoint hiccups. And it takes one run to check: point the model at a server that deliberately returns errors, and watch whether it reads the error and adjusts its arguments, or apologises and stops.
Muse Spark 1.2, briefly
Closed weights, API-only. 1,048,576-token context, ~131,072 output. Accepts text, images, video, audio and PDF.
For agent builders: structured output, parallel function calling, configurable reasoning effort, and explicit multi-agent support in either role โ primary planner or parallel subagent. It leans on planning, goal conditioning and context compaction to hold direction through long tasks.
Pricing is $1.25 / $4.25 per million tokens. Note the cost shape: with most models, cost tracks how much work the model does. With a 1M-context model, cost tracks how much context you chose to send โ a 200k-token prompt is $0.25 before it writes a word. Compaction is a feature for a reason.
What nobody has published
Meta has not documented the tool-calling wire format for either model. The model card and developer pages assert reliable tool use and precise schema invocation, but no serialization is specified. The reference runtimes simply expose an OpenAI-compatible surface.
So: routed through a provider, both models behave as OpenAI-shaped function callers, the same path every non-Anthropic model takes. Any claim you read about how Muse specifically handles deeply nested arguments is currently untested rather than documented.
Which makes your own server the only authority that counts.
Which one
- Glimmer when you want a model you can self-host, 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.
- Spark 1.2 when the task genuinely needs the context or parallel function calling โ whole-repository work, long debugging sessions, multi-agent orchestration.
The distillation relationship is the mental model: Glimmer inherits Spark's agentic behaviour in a package you can run at home; Spark keeps the scale, context and modalities that will not fit on a laptop.
You can run both Muse models against your own MCP server in the browser โ no install, no Meta account: Test your MCP server with Meta Muse
Full writeup with the complete benchmark tables: Muse Spark 1.2 and Muse Glimmer 30B for MCP
Sources: Meta AI Research โ Introducing Muse Glimmer ยท Scale AI MCP-Atlas leaderboard ยท Muse Glimmer on Hugging Face ยท Muse Spark 1.2 on OpenRouter
Top comments (0)