Size
Medium
30B
Tasks
93 scored · 7 unsupported (vision/audio)
Run date
Judge
TBD
Hardware
ASUS Ascent GX10
WikiBench 1.0
100 real tasks for models running locally. Each task scores 0–10. The overall score is a weighted mean, and partial credit counts. A separate model judges the answers: Ollama qwen2.5:7b.
Live rows from the public table. Cached for about a minute.
Size
Medium
30B
Tasks
93 scored · 7 unsupported (vision/audio)
Run date
Judge
TBD
Hardware
ASUS Ascent GX10
| Model | Overall | Size | Tasks | Run date | Judge | Hardware | Breakdown |
|---|---|---|---|---|---|---|---|
| NVIDIA Nemotron 3.5 Lightning 30B-A3BNVFP4Local (Spark)Medium | 5.03 | Medium30B | 93 scored · 7 unsupported (vision/audio) | TBD | ASUS Ascent GX10 |
Size
Spark-sized
122B
Tasks
93 scored · 7 unsupported (vision/audio)
Run date
Judge
TBD
Hardware
ASUS Ascent GX10
| Model | Overall | Size | Tasks | Run date | Judge | Hardware | Breakdown |
|---|---|---|---|---|---|---|---|
| Ornith 1.5 122BQ4Local (Spark)Spark-sized | 4.88 | Spark-sized122B | 93 scored · 7 unsupported (vision/audio) | TBD | ASUS Ascent GX10 |
Size
Large
552B
Tasks
100 scored · 0 unsupported
Run date
Judge
TBD
Hardware
Hosted (Nous Portal)
Size
Large
501B
Tasks
93 scored · 7 unsupported (vision/audio)
Run date
Judge
TBD
Hardware
Hosted (Nous Portal)
Size
Large
428B
Tasks
94 scored · 6 unsupported (vision/audio)
Run date
Judge
TBD
Hardware
Ollama cloud
| Model | Overall | Size | Tasks | Run date | Judge | Hardware | Breakdown |
|---|---|---|---|---|---|---|---|
| DeepSeek V4.1 FlashCloudLarge | 5.01 | Large552B | 100 scored · 0 unsupported | TBD | Hosted (Nous Portal) | ||
| Reflection Beam 1.0 501B-A23BCloudLarge | 4.73 | Large501B | 93 scored · 7 unsupported (vision/audio) | TBD | Hosted (Nous Portal) | ||
| MiniMax M3.1 FlashCloudLarge | 4.29 | Large428B | 94 scored · 6 unsupported (vision/audio) | TBD | Ollama cloud |
WikiBench 1.0 is the first live version of this benchmark. It has 100 tasks in 12 weighted categories: agentic, tool use, computation, coding, visualization, security, research/writing, hallucination, humanized, computer use, vision, transcription.
Size classes are Tiny (4B parameters or fewer), Small (5–9B), Medium (10–27B), and Large (a model that fits on one DGX Spark). If a run does not report a class, the row stays TBD. Parameter counts are shown only when the run reported them.
Unsupported tasks are counted and left out of the mean. They are not scored as zero. If a run has not reported category scores, the breakdown says pending. This page does not fill those in.
The Spark runner posts each finished result file to a token-protected endpoint with scripts/wikibench_post_result.py. New results land as drafts in wikibench_results and appear here only after they are reviewed and published. The public key can only read. Site research notes live on the methodology page.