Nemotron-3.5-Bolt + dflash + mmproj

Quality & speed (tokens/s) — LLM-judge scores 0–1; cells show mean ± standard error (SE)

Statistics: each cell is the MEAN (arithmetic average of scored samples) ± STANDARD ERROR of the mean (SE = sample standard deviation / √n); (n/N) = scored/total samples. Scores are LLM-judge quality 0–1 (higher = better); TPS = output tokens/second (higher = faster).

Quality / ratings

Image understanding

Metric	Score (0–1)
Analysis	0.82 ± 0.030
Total question-answer	0.62 ± 0.22
Public transit timetable reading	0.56 ± 0.20

Summarization

Metric	Score (0–1)
English	0.73 ± 0.026

Translation

Combined (both directions)

Metric	Score (0–1)
English (all en→X)	0.67 ± 0.018
African / Middle eastern	0.73 ± 0.035
Western European	0.70 ± 0.017
Eastern European	0.69 ± 0.037
Asian	0.66 ± 0.028
All pairs	0.69

To English (X→en)

Metric	Score (0–1)
African / Middle eastern	0.72 ± 0.052
Western European	0.69 ± 0.025
Eastern European	0.76 ± 0.048
Asian	0.73 ± 0.049
All X→en	0.71 ± 0.019

From English (en→X)

Metric	Score (0–1)
African / Middle eastern	0.75 ± 0.022
Western European	0.71 ± 0.024
Eastern European	0.62 ± 0.049
Asian	0.64 ± 0.032
All en→X	0.67 ± 0.018

Audio transcription

Metric	Score (0–1)
Short (1 min)	0.94 ± 0.060
Medium (5 min)	0.73 ± 0.033
Long (30 min)	0.50 ± 0.20

Tokens per second (TPS) — average: 180