Using the Meta Model API, we at Box put Muse Spark 1.1 through the Box Complex Work Eval, our benchmark of realistic, document-grounded tasks drawn from the analytical work that fills actual enterprise workflows. The model reads a set of source documents, reasons across them, and produces a finished deliverable: an analysis, a report, a due-diligence verdict, a review. It’s the kind of long-horizon, multi-step work where a model has to stay accurate over many moves, not just answer a single question. Here’s what we found.
The short version: Muse Spark 1.1 holds its own against the best models available today.
Across the bulk of the benchmark, it performs at or above today’s models and on a number of tasks it’s the standout — turning messy, multi-source data into the right answer where other leading models come up short. It's a genuinely capable model for enterprise knowledge work.
Consistent across the breadth of enterprise work
The strongest signal in the results isn’t any single score, but rather, the breadth. Muse Spark 1.1 keeps pace with the leading models across most of the industries we test, and it’s especially at home in structured, procedural work: the professional services, public sector, and industrial tasks where the job is to follow a defined process over a stack of documents and get every step right.
In those categories, it pulls ahead of the top-tier composite by as much as 5 to 6 points. This is the connective tissue of enterprise work — intake, review, reconciliation, reporting — and it’s where Muse Spark 1.1 is most consistently hitting the mark.
Muse Spark 1.1 is also strong on report drafting. Given raw inputs and a format to fill, it produces clean, complete, well-organized deliverables that hold up against what the best models generate. For the day-to-day work of turning data and documents into something a person can act on, Muse Spark 1.1 is a dependable choice.
Where it shines: Turning data into the right answer
The most striking results came on data-analysis tasks — reading imperfect, real-world source data and computing a defensible answer. On several of these, Muse Spark 1.1 didn’t just match the pack, it clearly outperformed it. A few examples:
- A cost-optimization analysis: Given operational data spread across a workbook, the task was to derive a set of efficiency metrics per market and recommend where to focus. Muse Spark 1.1 carried the multi-step calculation through cleanly and reached the correct prioritization, outperforming the high-performing foundational models by nearly 30 points where other models stumbled on the intermediate math.
- A client-portfolio analysis:The task required segmenting a set of accounts, computing the right splits, and identifying the standout client on the correct dimension with a tempting wrong dimension sitting right next to it in the data. Muse Spark 1.1 grounded its answer in the right column and got it right, beating the composite by more than 25 points; a common failure here is anchoring on the nearby distractor.
- A large cost-rollup: Reconciling planned versus actual costs across a long itemized maintenance record, the task asked for several totals and variances to tight tolerances. Muse Spark 1.1 produced figures that tied out across the board. This is the kind of careful, additive work that rewards not cutting corners.
The throughline is that Muse Spark 1.1 is at its best when the work is quantitative but legible — when there's a correct computation to be done over real data and the model has to resist the easy approximation. On that class of task, it’s often ahead of state-of-the-art models.
Where Muse Spark 1.1 is headed
No model is uniformly ahead, and Muse Spark 1.1 is no exception. The hardest, densest multi-step reasoning is where any model earns its keep, and it’s the natural place to keep pushing. But the headline is how much of the leading-edge this model already reaches. Across the breadth of real enterprise work (the structured processes, the report drafting, and especially the data analysis where getting the number right is the whole job) Muse Spark 1.1 performs like a model built for the work enterprises actually do.
