← All models
Proving Ground · Rank 3 of 6
Gemma 4 31B
Google's 31B dense model — the top of the Gemma 4 size curve on one desktop-sized AI computer.
94%
517 of 550 graded answers correct
550 tasks × 3 passes
Setup and speed
Model
Gemma 4 31B
Google's 31B dense model — the top of the Gemma 4 size curve on one desktop-sized AI computer.
Quantization
QAT w4a16 compressed-tensors (4-bit weights, 16-bit activations)
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 vLLM 0.29.0. One desktop-sized AI computer runs the whole model.
Source
Google DeepMind
Model by Google DeepMind — the QAT (quantization-aware trained) checkpoint, served with vLLM's native Gemma 4 tool-call parser so tool use is the model's own, not a harness shim.
First token
0.12s
starts responding almost instantly
Writing speed
9.6 w/s
the pace of the prose once it starts
Tokens per second
10.8
the raw generation rate behind that
Visible answer starts
immediately
answers straight away — no separate reasoning phase
Typical task
18.9s
half finish faster than this
Slow tasks
76.0s
9 in 10 finish inside this
Full results
Score by task difficulty
easy
98.1%
One clear step — read the request, produce the answer.
573 graded answers
medium
95.4%
Several steps or a judgment call: reconcile records, pick a queue, pull a field.
714 graded answers
hard
84.6%
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
95%
95/100 correcttypical 3.1s
Document extraction
91%
91/100 correcttypical 37.3s
Spreadsheet analysis
98%
98/100 correcttypical 81.8s
Formulas and SQL
86%
86/100 correcttypical 14.0s
Business writing
97%
97/100 correcttypical 18.9s
Classification and routing
100%
50/50 correcttypical 2.3s
Where it missed
- —Each task is run three times; the score counts all 1,650 attempts, not 550 tasks.
- —509 of 550 tasks were correct on all three passes. A perfect 50/50 on classification and routing, and 100% on both easy and medium spreadsheet tasks.
- —The single weak workload is Formulas & SQL at 86, where the hard tier drops to 46.7 — the same place every model in this suite struggles.
- —Nothing was cut off by the output limit.
The first token arrives at 0.12s and the answer starts straight away — this model does not run a separate reasoning phase. The 18.9s typical task time is the answer being written.
Files it produced
This model has not run the file round yet. When it does, its annual report, workbook, investor deck and social posts will be published here, unedited, with their scores.