Meta: Muse Spark 1.2

Analyze this quarterly sales spreadsheet and identify the three product categories with declining revenue. Explain what might be causing each drop and suggest one tactical fix per category.

In this illustrative example, Muse Spark would parse the uploaded file, identify trends across columns, and return a structured analysis. It might note that Category A saw a 12% decline correlated with increased competitor pricing pressure, Category B's drop aligns with seasonal patterns from prior years, and Category C's revenue fell after a key retailer reduced shelf space. For each, it would propose a concrete action: a targeted discount campaign for A, inventory adjustment for B's known cycle, and a direct outreach program to the retailer for C. The response would reference specific cells and time periods from the spreadsheet.

With a 1M+ token context window, Muse Spark can ingest large files without summarization loss—useful for multi-tab spreadsheets or dense reports. This example shows its multimodal file handling and analytical reasoning. The trade-off: at $4.25/Mtok output, verbose responses on complex datasets get expensive quickly compared to text-only models.

I'm attaching a 90-second product demo video. Write a two-paragraph marketing description that captures the key features shown, then suggest three social media captions optimized for LinkedIn, Twitter, and Instagram.

In this illustrative example, Muse Spark would process the video frames and audio, then generate a polished marketing description highlighting the product's interface, core workflow, and differentiators visible in the demo. The LinkedIn caption might emphasize ROI and enterprise use cases in a professional tone, the Twitter version would be punchy with a clear CTA and relevant hashtags, and the Instagram caption would lead with an emotional hook and include emoji for platform fit. Each caption would reference specific moments from the video timestamp.

Video understanding sets Muse Spark apart from text-only LLMs—it can extract visual context that a transcript alone would miss, like UI polish or user reactions. This matters for content teams working from raw footage. The caveat: without public benchmarks, it's unclear how its video comprehension compares to GPT-4o or Gemini 1.5 Pro on complex scenes.

I'm pasting a 40-page legal contract. Highlight any clauses that create unilateral termination rights for the vendor, then draft a one-page memo for our CFO explaining the financial exposure and recommending two contract amendments.

In this illustrative example, Muse Spark would scan the full contract within its 1M+ token window, flag sections 8.3, 12.1, and Appendix C as containing unilateral termination language, and quote the exact wording. The memo would quantify exposure—e.g., "Clause 8.3 allows termination with 30 days' notice, putting $2.4M in annual licensing at risk"—and propose amendments like mutual termination rights or a 180-day notice period. The tone would be executive-ready: concise, risk-focused, and actionable.

Long-context models like Muse Spark excel at document review tasks that would require chunking or RAG pipelines in smaller models. The 1M token window handles most contracts in one pass. The risk: at $1.25/Mtok input, repeatedly uploading large documents for iterative edits costs more than models with cheaper input pricing, so batch your questions.

When you need one model to handle decks, videos, and contracts

A 4-person consulting shop sends client deliverables through Switchy before final review—PowerPoint decks, recorded Loom walkthroughs, PDF contracts, and Notion briefs. Meta Muse Spark handles all four modalities in a single 1M-token context window, so you can drop an entire project folder and ask "flag any inconsistencies between the deck narrative and the contract terms." At $1.25 input per Mtok, a 200k-token mixed review costs $0.25. The trade-off: no public benchmarks yet, so you're trusting Meta's internal evals. If your deliverables are text-only or you need proven accuracy on specialized tasks, wait for benchmark data or test another model first. For teams already juggling four tools to review four formats, this consolidates the workflow into one pass.

How a 10-person trust-and-safety team uses video understanding at scale

A social platform's moderation queue includes 90-minute livestream recordings flagged by users. The team needs frame-level analysis plus transcript context to decide on strikes. Muse Spark's video modality and 1M-token window let you ingest the full stream, ask "identify segments with policy violations and timestamp them," and get structured output without chunking. At $4.25 output per Mtok, a 50k-token moderation report costs $0.21. The model's multi-format input means you can also attach the platform's community guidelines PDF in the same prompt for policy-grounded decisions. The risk: absent public benchmarks, you'll want to A/B test against a proven vision model on your first 500 flags before going all-in. If accuracy holds, this is the cheapest way to handle video moderation without a preprocessing pipeline.

When M&A analysts need to compare 800-page filings in one query

A 6-person investment team runs due diligence on acquisition targets. They upload 10-Ks, investor decks, and legal exhibits—often 600k tokens combined—and ask Muse Spark to surface discrepancies between reported revenue and footnote adjustments. The 1M-token context means no document gets split across prompts, preserving cross-reference accuracy. Input cost is $0.75 for a 600k-token diligence run; if the output summary is 30k tokens, you pay $0.13 more. Total per deal: under $1. The catch: without MMLU-Pro or FinanceBench scores, you're flying blind on numerical reasoning quality. Run a parallel check with a benchmarked model on your first 10 deals. If Muse Spark's answers align, you've found a multi-format diligence tool that costs 60% less than alternatives with similar context windows.