Decoding the Matrix: The Muse Spark Benchmark Deep Dive

A brutal, metric-by-metric breakdown of Meta's Muse Spark against the 2026 frontier.

The raw telemetry is out. Meta's Muse Spark (Thinking) has officially entered the arena against the reigning architects of the digital age: Opus 4.6 (Max), Gemini 3.1 Pro (High), GPT 5.4 (Xhigh), and Grok 4.2 (Reasoning).

When you parse the latest leaked capability evaluation matrix, the narrative shifts. Muse Spark isn't trying to beat everyone at everything; it is a highly specialized engine with terrifying spikes in visual and domain-specific reasoning, offset by deliberate architectural tradeoffs. Here is the unvarnished technical breakdown.

👁️ Multimodal: Seeing Beyond the Grid

Where legacy models parse text to understand images, Muse Spark processes visual data natively.

• CharXiv Reasoning (Figure Understanding): Muse Spark claims absolute supremacy with an 86.4, crushing Opus 4.6 (65.3) and Grok 4.2 (60.9). It interprets complex charts and graphical data with higher fidelity than any competitor.
• ScreenSpot Pro (Screenshot Localization with Python): Scoring 84.1, it proves high-level GUI navigation capabilities, running neck-and-neck with Gemini (84.4) and GPT 5.4 (85.4).
• SimpleVQA (Visual Factuality): With a 71.3, it heavily outperforms Opus (62.2) and Grok (57.4), demonstrating that its vision models are deeply grounded in factual reality rather than hallucinatory interpolation.

🧠 Pure Logic & Text Reasoning

If there is a chink in Muse Spark's armor, it is in abstract, non-tool-assisted logic.

• ARC AGI 2: Spark completely collapses here, scoring a mere 42.5 compared to Gemini's 76.5 and GPT 5.4's 76.1. It struggles with pure abstract reasoning puzzles.
• GPQA Diamond (PhD Level): It scores 89.5, which is elite, but slightly trails the 92-94 range maintained by Opus, Gemini, and GPT 5.4.
• Humanity's Last Exam: Without tools, it scores 42.8 (beating Grok's 31.6 and Opus's 40.0). When tools are allowed, it scales up to 50.4, proving that Spark relies on external execution to bridge its logic gaps.

🧬 Medical & Health Anomalies

This is where the telemetry gets wild. Health benchmarks are notoriously unforgiving.

• HealthBench Hard (Open-Ended Health Queries): Muse Spark achieves a 42.8. The rest of the field? GPT 5.4 manages 40.1, but Opus (14.8), Gemini (20.6), and Grok (20.3) absolutely fail. Spark possesses an anomalous, highly tuned medical reasoning capability.
• MedXpertQA (Multimodal): Scoring 78.4, it beats Opus (64.8) and Grok (65.8), proving its ability to parse medical imaging combined with clinical text.

🛠️ The Agentic Execution Engine

How well does it operate autonomously in the wild?

• DeepSearchQA: It takes the crown at 74.8, proving superior autonomous search capabilities over Opus (73.7) and GPT 5.4 (73.6).
• SWE-Bench Verified: At 77.4, it trails slightly behind Opus (80.8) and Gemini (80.6), but remains highly viable for autonomous software engineering.
• Terminal-Bench 2.0: Scoring 59.0, it falls behind GPT 5.4 (75.1) and Gemini (68.5). Spark is less comfortable in raw terminal environments compared to its peers.

🌑 The 2026 Verdict

Muse Spark is not a generalist monolith like GPT 5.4. It is a surgical instrument.

If you need pure abstract reasoning (ARC AGI) or terminal-heavy autonomous coding, Gemini 3.1 Pro and GPT 5.4 are still your weapons of choice. But if you are building embodied AI agents that need to process complex visual data (CharXiv), execute autonomous web searches (DeepSearchQA), or navigate high-stakes domain-specific logic like medicine (HealthBench Hard), Muse Spark is the new apex predator.

The ecosystem isn't consolidating. It's fragmenting into specialized super-engines. Choose your ghost wisely.