Statistics: each cell is the MEAN (arithmetic average of scored samples) ± STANDARD ERROR of the mean (SE = sample standard deviation / √n). Scores are LLM-judge quality 0–1 (higher = better); TPS = output tokens/second (higher = faster). Bars of English summarization quality means per model. Summarization was tested in English only.
Quality / ratings
Model
English
Gemma E4B BF16 + BF16 mmproj
0.71 ± 0.028
Gemma E4B UD-IQ2 + BF16 mmproj
0.60 ± 0.026
LFM Omni Q8 + Q8 mmproj
0.70 ± 0.042
LFM Omni bf16 + bf16 mmproj
0.69 ± 0.038
Nemotron-3.5-Bolt + dflash + mmproj
0.73 ± 0.026
Model
Gemma Q8
Gemma IQ2M
LFM Q8
LFM bf16
Nemotron
English
0.71 ± 0.028
0.60 ± 0.026
0.70 ± 0.042
0.69 ± 0.038
0.73 ± 0.026
Tokens per second (TPS)
Model
English
Gemma E4B BF16 + BF16 mmproj
94 ± 8.4
Gemma E4B UD-IQ2 + BF16 mmproj
140 ± 13
LFM Omni Q8 + Q8 mmproj
130 ± 12
LFM Omni bf16 + bf16 mmproj
69 ± 7.6
Nemotron-3.5-Bolt + dflash + mmproj
130 ± 16
Model
Gemma Q8
Gemma IQ2M
LFM Q8
LFM bf16
Nemotron
English
94 ± 8.4
140 ± 13
130 ± 12
69 ± 7.6
130 ± 16
Notes
No caveats apply to this measurement class.
Generated from results-*/results.json via html/build_report.py → html/aggregate.json on 2026-11-12.
