niklaslenzGemma 4 31B in the Gauntlet
1 DGX Spark 128 GB · TensorFold 1.2.5 (official release), Ash's recipe (Gemma 4 31B, MLX 4-bit, DFlash drafter)
What happened on the course
Each range is one task with one answer and no retry. Every wrong or missing target is a miss and costs a penalty loop. The race time is thinking, running, shooting and loops together.
| # | Task | Targets | Hits | Thinking | Buzzer | Penalty loops | Range time |
|---|---|---|---|---|---|---|---|
| 1 | River ferry (puzzle, level 1) | ✕✕✕✕✕ no answer in time | 0 / 5 | 35.0 s | buzzer, answer forced | 5 (60.0 s) | 102.2 s |
| 2 | Painter (code, level 2) | ●●●✕✕ | 3 / 5 | 8.8 s | no | 2 (24.0 s) | 40.0 s |
| 3 | Bench seating (puzzle, level 3) | ●●●●● | 5 / 5 | 13.8 s | no | 0 (0.0 s) | 21.0 s |
| 4 | Snake (code, level 4) | ●✕●✕● | 3 / 5 | 21.9 s | no | 2 (30.0 s) | 59.1 s |
● hit, ✕ miss. The range time runs from arriving at the mat to leaving the range, loops included. A buzzer means the thinking cap ran out and the game forced the answer.
Every number has its file
Result (result.json) Race log (log.jsonl) Model answers (io.jsonl) System prompt Exact command Verify output Model config
Version: prompt version gauntlet-4, tasks bia-2, rules 1, prompt fingerprint 07fff686, tasks fingerprint c2c0f4f6, forced-answer style notes. Verified: the replay of the race log reproduces the end state and every code answer re-runs to the same verdict. Forced-answer style: how the game asks for the answer when the buzzer sounds (a prefilled answer start, or notes when the engine refuses that). Command and config are normalized: internal addresses appear as 127.0.0.1.
Who made what
- Recipe (docs/recipes/gemma-4-native.md)Ash HartX @ashxhartGitHubHugging FaceWebsite
- Engine (TensorFold 1.2.5, official release)Ash HartX @ashxhartGitHubHugging FaceWebsite
- Model (Gemma 4 31B, built by Google DeepMind)GoogleHugging FaceWebsite
- MLX 4-bit weights (revision 696d436c, converted from Google's checkpoint)MLX CommunityHugging Face
- Draft model (DFlash, revision eabd6483)z-lab (Zhijian Liu)X @zhijianliu_Hugging FaceGitHub
Full v2 results of the matching setup (same model, engine, quant and hardware; the engine build of the race is not recorded in the race files) · How the game test works · All races