← All models
Proving Ground · Rank 4 of 6

Nemotron 3.5 Bolt 74B-A9B

NVIDIA's 74B mixture-of-experts, packed to 24GB so it fits on one desktop-sized AI computer.

91.5%
503 of 550 graded answers correct
550 tasks × 3 passes

Setup and speed

Model
Nemotron 3.5 Bolt 74B-A9B
NVIDIA's 74B mixture-of-experts, packed to 24GB so it fits on one desktop-sized AI computer.
Quantization
IQ2_XXS experts + NVFP4 Mamba + Q8_0 attention (24GB pack)
Compressed so it fits the hardware below — a higher or lower bit-width version would score differently
Deployment
1× DGX Spark (dgx-spark-x1)
Served by TensorFold v1.0.12 (libbolt). One desktop-sized AI computer runs the whole model.
Source
Mia
Serving recipe by Mia — NVIDIA's Nemotron 3.5 Bolt 74B-A9B 24GB pack on TensorFold's Zig engine, with a LiLiCorr drafter and a Bolt backend that puts the whole model on one DGX Spark.
First token
0.06s
starts responding almost instantly
Writing speed
71 w/s
the pace of the prose once it starts
Tokens per second
71
the raw generation rate behind that
Visible answer starts
immediately
answers straight away — no separate reasoning phase
Typical task
27.3s
half finish faster than this
Slow tasks
92.3s
9 in 10 finish inside this

Full results

Score by task difficulty
easy
94.2%
One clear step — read the request, produce the answer.
573 graded answers
medium
90.9%
Several steps or a judgment call: reconcile records, pick a queue, pull a field.
714 graded answers
hard
89%
Multi-step work with a trap: a policy to apply, a dataset to analyze, a budget to stay inside.
363 graded answers
Score by kind of work
Inbox triage and reply
99%
99/100 correcttypical 19.6s
Document extraction
98%
98/100 correcttypical 24.7s
Spreadsheet analysis
95%
95/100 correcttypical 33.9s
Formulas and SQL
86%
86/100 correcttypical 35.4s
Business writing
75%
75/100 correcttypical 45.8s
Classification and routing
100%
50/50 correcttypical 9.2s
Where it missed
  • —Each task is run three times; the score counts all 1,650 attempts, not 550 tasks.
  • —503 of 550 tasks were correct on all three passes. The losses concentrate in two places: business writing and Formulas & SQL.
  • —71 business-writing replies were real newlines inside a JSON string instead of escaped ones, so the harness could not parse them. 62 of those parse cleanly once escaped and carry every required field — the writing is substantively right and the formatting is wrong.
  • —Formulas & SQL is the genuine capability limit: 13 of 15 failing tasks miss on all three passes, inventing schema that does not exist and writing range formulas that do not evaluate.
  • —Three attempts on one inbox task burned the full output budget while reasoning and returned nothing — reported as truncated, not as a wrong answer.
  • —Finale: three of four artifacts pass. The deck initially produced nothing at all with reasoning on — the model spent the entire output budget re-reading files it already had. Re-run with reasoning off, it built a complete ten-slide deck; that setting is recorded on this row.

The first token arrives at 0.06s and the answer starts straight away — this model does not run a separate reasoning phase. The 27.3s typical task time is the answer being written.

Files it produced

Four real documents from one fact sheet, produced once. Passed is machine-checked against the fact sheet and the brief; quality is a separate 1–5 score from a published rubric and never changes whether a file passed. How files are graded →

Passed 3/4Quality 2.50 / 5
Annual reportquality 2.00

An 8–12 page annual report with six required sections: shareholder letter, financial review, category performance, customer and geographic performance, risk factors, and outlook.

Download DOCX
Supporting workbookquality 2.75

A two-sheet workbook: a Summary comparing FY2025 to FY2024, and a Monthly sheet with all twelve months.

Download XLSX
Investor deckquality 3.00

A ten-slide investor presentation following a fixed outline, with the brand palette applied and at least one real chart object.

Download PPTX
Social announcementsquality 2.25

An X post of 280 characters or fewer and a LinkedIn post, both announcing the results, text only.

Download MD

Every artifact below is the file the model actually produced, published unedited so you can open it yourself. These are AI-generated documents about a fictional company — read them as samples of the model's work, not as advice or a real annual report. The workbook's formula cells hold real formulas — a viewer that does not recalculate may show them blank until Excel opens the file.