Compare runs
Pick two or more runs to compare. Gemma-4-31B on M4 Max 128 GB is the first; add another below.
| Run | Gemma-4-31BM4 Max 128 GB · llama.cpp llama-swap (version not exposed)Take out Gemma-4-31B |
|---|---|
| Score | —No score |
| Tasks passed | 27 of 44 |
| Points | 43 of 64 |
| Speed | 9.8 tok/s |
| Tokens | 318,255 |
| Summarizing | 0/5 |
| One-sentence summary | Not run |
| Meeting notes in three bullets | Not run |
| Review sentiment and complaint | Not run |
| Faithful summary of a study | Not run |
| Action items from an email thread | Not run |
| Extracting | 0/5 |
| Fields from an email as JSON | Not run |
| Every email address, once | Not run |
| A list as CSV | Not run |
| The real deadline | Not run |
| Warnings and errors from a log | Not run |
| Following instructions | 0/5 |
| Exactly one word | Not run |
| Three lowercase lines | Not run |
| Exact JSON shape | Not run |
| Sort a list | Not run |
| A name with three rules | Not run |
| Reasoning | 5/5 |
| Unit price | Passed13.4 tok/s |
| Days of the week | Passed13.4 tok/s |
| Who owns the cat | Passed13.5 tok/s |
| Arrival time | Passed13.5 tok/s |
| Filling a tank | Passed13.4 tok/s |
| Reading code | 4/4 |
| What does it print | Passed13.4 tok/s |
| Trace a loop | Passed13.3 tok/s |
| Find the bug | Passed13.4 tok/s |
| Run a query in your head | Passed13.5 tok/s |
| Using tools | 5/5 |
| Call one tool | Passed13.4 tok/s |
| Don't call a tool when none is needed | Passed13.5 tok/s |
| Two calls for two answers | Passed13.5 tok/s |
| Chain two tools | Passed13.5 tok/s |
| Handle a failed call honestly | Passed13.5 tok/s |
| Building pages | 5/5 |
| A small website | Passed13.3 tok/s |
| A pricing table with a toggle | Passed13.3 tok/s |
| An accessible sign-up form | Passed13.3 tok/s |
| A to-do app that remembers | Passed13.3 tok/s |
| A bar chart in SVG | Passed13.3 tok/s |
| Building games | 1/2 |
| Build 2048 | Passed12.9 tok/s |
| Build Snake | Failed9.4 tok/s |
| Writing code | 2/3 |
| Merge intervals | Passed13.3 tok/s |
| An LRU cache | Passed13.3 tok/s |
| Evaluate arithmetic | Failed9.4 tok/s |
| Fixing code | 2/2 |
| Fix the bookings module | Passed13.2 tok/s |
| Fix the CSV reader | Passed13.2 tok/s |
| Planning and data | 3/3 |
| A query over hidden data | Passed13.3 tok/s |
| Schedule five meetings | Passed13.1 tok/s |
| A replacement in ten calls or fewer | Passed13.2 tok/s |
Scrolls sideways to show every run.
A check marks the best of each row. tok/s is tokens a second: how fast the model writes. How the score works.