Granite Guardian 4.1 8B · file← Granite Guardian 4.1 8B
MLX 4-bit g64
A 4-bit MLX copy for a speed test. Never scored; deleted. Scratch build, made here.
- Size
- 4.4 GB
- Status
- Deleted 16 Sep 2026
- Scored runs
- 0
ResultsNever scored
This file never completed a scored run. Its record below says why.
The recordHow it was made · how it was served · what went wrong · what happened to it
Rebuilt from the lab's records by Claude, an AI model. The results above come straight from the automated runs. How these pages are made.
| File | _mlxtest/granite-guardian-4.1-8b-mlx-4bit/ |
| Format | MLX 4-bit, group 64, 4.4 GB by du |
| Where from | made here 2026-09-11 ~13:17Z by rapid-mlx's mlx_lm |
| Status | deleted 2026-09-16 20:10Z |
How it was made
- Input: the safetensors in the HF cache (Granite Guardian 4.1 8B BF16 safetensors (IBM, HF cache)).
- Command:
MPY=.local/share/uv/tools/rapid-mlx/bin/python
$MPY -m mlx_lm convert --hf-path ibm-granite/granite-guardian-4.1-8b \
--mlx-path models/_mlxtest/granite-guardian-4.1-8b-mlx-4bit -q --q-bits 4 --q-group-size 64- Tool: mlx_lm inside the rapid-mlx uv tool env (version not printed; the run log lists mlx_lm 0.31.3 for rapid-mlx as of 2026-10-21).
- Date: 2026-09-11 ~13:17–13:18Z (local 14:17–14:18). Size: 4.4 GB by du.
- Why: to compare decode speed of MLX 4-bit against GGUF Q4_0 rather than Q4_K_M. Alan had asked to settle the MLX comparison (reconstructed).
How it was served
| label | dates (UTC) | binary and version | flags | harness profile | why this label exists |
|---|---|---|---|---|---|
| — (no run-log label) | 2026-09-11 13:18Z | mlx_lm / rapid-mlx | — | — | speed probe only |
Decisions
- Built for a speed comparison the model could not support (a judge emits ~11 tokens).
- Deleted 2026-09-16 20:10Z with
_mlxtest/("SPENT — result already in findings") on Alan's "Go for it then". Never screened.
What went wrong
- Void speed probe: tokens per second was computed from requested tokens, not generated tokens.
What happened to it
Converted 2026-09-11 → void speed probe → deleted 2026-09-16 (4.4 GB). No RESTORE entry. Re-creatable from the HF-cache safetensors.