# Muse Spark

Muse Spark is an AI model released by Meta on Apr 8 2026. Tracked results include 89.5% on GPQA Diamond, 77.4% on SWE-Bench Verified and 55% on SWE-Bench Pro.

## Facts

| Field | Value |
| --- | --- |
| Model | Muse Spark |
| Developer | Meta |
| Release date | Wednesday, Apr 8 2026 |
| Licensing | Closed |

## Tracked benchmark scores

| Benchmark | Score | Source | What it measures |
| --- | --- | --- | --- |
| Gray Swan IPI (k = 1) | 2.9% | Lab | Attackers hide malicious instructions inside content the AI reads — a web page, an email, a document — and try to hijack what it does. Gray Swan's indirect prompt injection benchmark measures how often such an attack succeeds when the attacker gets a single try. Lower is better. |
| Gray Swan IPI (k = 10) | 14.3% | Lab | Attackers hide malicious instructions inside content the AI reads — a web page, an email, a document — and try to hijack what it does. This variant gives the attacker 10 tries and counts an attack as successful if any of them works. Lower is better. |
| Gray Swan IPI (k = 15) | 16.5% | Lab | Attackers hide malicious instructions inside content the AI reads — a web page, an email, a document — and try to hijack what it does. This variant gives the attacker 15 tries and counts an attack as successful if any of them works. Lower is better. |
| SWE-Bench Pro | 55% | Lab | Can the AI fix real bugs in real software? It's handed actual problems from open-source projects and has to write code that genuinely solves them. Higher is better. |
| SWE-Bench Verified | 77.4% | Lab | Real coding tasks pulled from open-source projects — the AI has to find and fix actual bugs. A human-checked version of the original SWE-Bench. Higher is better. |
| DeepSWE 1.1 | 10% | Lab | Artificial Analysis' independent test of deep, agentic software-engineering work — the AI has to plan and carry out substantial coding tasks end to end. (Version 1.1 of the test.) Higher is better. |
| Terminal-Bench 2.1 | 67.3% | Lab | Can the AI work in a command-line terminal — running commands and finishing technical setup tasks the way a developer would? Higher is better. |
| MCP Atlas | 82.2% | Lab | Can the AI chain together many tools and steps to complete one bigger task, rather than doing just a single thing? Higher is better. |
| JobBench | 17% | Lab | Tests the AI on professional workplace tasks that require using real work tools — the kind of multi-step jobs an office worker handles. Higher is better. |
| Toolathlon-Verified | 49.4% | Lab | Tests how well the AI uses everyday personal tools and apps to get things done — a human-checked version of Toolathlon. Higher is better. |
| Humanity's Last Exam (with tools) | 50.4% | Lab | Humanity's Last Exam — extremely hard expert questions across many subjects. “With tools” means the AI is allowed to search the web or run code while answering. Higher is better. |
| ARC-AGI-2 | 42.5% | [BenchLM](https://benchlm.ai), retrieved 2026-07-11 | Puzzle-style tests of abstract reasoning and pattern-finding — the kind of thing people find easy but AIs often struggle with. Higher is better. |
| GPQA Diamond | 89.5% | Lab | Graduate-level science questions in biology, physics, and chemistry — hard enough that subject-matter PhDs score around 65%. Higher is better. |
| OSWorld-Verified | 53.3% | Lab | Can the AI actually operate a computer — clicking, typing, and using real apps — to finish tasks on its own? Higher is better. |
| CharXiv Reasoning | 88.9% | Lab | Can the AI read and reason about complex charts and figures, not just text? Higher is better. |
| BabyVision | 39.9% | Lab | Tests core visual reasoning — seeing and understanding images the way even young children can, which AIs often find surprisingly hard. Higher is better. |
| MMMU-Pro | 80.4% | Lab | A tougher version of MMMU — college-level questions that mix images, diagrams, and text together. Higher is better. |

## About Muse Spark

Muse Spark was introduced on April 8, 2026 as the first Muse-family model from Meta Superintelligence Labs. Meta described it as a natively multimodal reasoning model with tool use, visual reasoning and multi-agent orchestration. It launched in Meta AI, with a private API preview for selected users.

Meta presented the release as the result of rebuilding its model-training stack, with further scaling planned. The announcement also described Contemplating mode, which used multiple agents reasoning in parallel and was scheduled for gradual rollout. That product mode was distinct from an ordinary single-model response, and the launch did not establish that Meta had permanently retired the Llama name.

## Questions and answers

### When was Muse Spark released?

Muse Spark was released by Meta on Wednesday, Apr 8 2026.

### Who made Muse Spark?

Muse Spark was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research.

### What benchmark scores did Muse Spark get?

Muse Spark reports 17 tracked benchmark scores — Gray Swan IPI (k = 1): 2.9%; Gray Swan IPI (k = 10): 14.3%; Gray Swan IPI (k = 15): 16.5%; SWE-Bench Pro: 55%; SWE-Bench Verified: 77.4%; DeepSWE 1.1: 10%; Terminal-Bench 2.1: 67.3%; MCP Atlas: 82.2%; JobBench: 17%; Toolathlon-Verified: 49.4%; Humanity's Last Exam (with tools): 50.4%; ARC-AGI-2: 42.5%; GPQA Diamond: 89.5%; OSWorld-Verified: 53.3%; CharXiv Reasoning: 88.9%; BabyVision: 39.9%; MMMU-Pro: 80.4%. Tracked scores may come from lab reports or independent benchmarks; source details accompany the benchmark data.

### Is Muse Spark open source?

No. Muse Spark is a closed model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms.

### What came before and after Muse Spark?

Meta's previous tracked release was LLaMA 4 Maverick on Apr 5 2025, 368 days earlier. It was followed by Muse Spark 1.1 on Jul 9 2026.


---

Canonical page: https://aireleasetracker.com/model/meta/muse-spark
Site index: https://aireleasetracker.com/llms.txt
Source: AI Release Tracker (https://aireleasetracker.com). Most benchmark scores come from lab launch material; gathered results identify the leaderboard that published them.
