LFM Omni bf16 + bf16 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	0.72 ± 0.041
Total question-answer	0.40 ± 0.24
Public transit timetable reading	0.36 ± 0.22

Summarization

Metric	Score (0–1)
English	0.69 ± 0.038

Translation

Combined (both directions)

Metric	Score (0–1)
English (all en→X)	0.37 ± 0.027
African / Middle eastern	0.37 ± 0.063
Western European	0.49 ± 0.029
Eastern European	0.32 ± 0.042
Asian	0.30 ± 0.030
All pairs	0.40

To English (X→en)

Metric	Score (0–1)
African / Middle eastern	0.43 ± 0.075
Western European	0.50 ± 0.039
Eastern European	0.27 ± 0.060
Asian	0.35 ± 0.061
All X→en	0.43 ± 0.029

From English (en→X)

Metric	Score (0–1)
African / Middle eastern	0.25 ± 0.11
Western European	0.48 ± 0.045
Eastern European	0.38 ± 0.055
Asian	0.28 ± 0.033
All en→X	0.37 ± 0.027

Audio transcription

Metric	Score (0–1)
Short (1 min)	0.88 ± 0.058
Medium (5 min)	0.78 ± 0.095
Long (30 min)	no data

Tokens per second (TPS) — average: 93