Granite Guardian 4.1 8B · file← Granite Guardian 4.1 8B
MLX 8-bit
A larger MLX copy, made to find what went wrong. Deleted the same day. Made here.
- Size
- 8.3 GB
- Status
- Deleted 9 Sep 2026
- Scored runs
- 1
- Smallest Mac
- 16 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| screeninggranite-guardian-4.1-8b-mlx-q8 | 98.3% | 91.1–99.7 | Passed 59/60 | mlx_lmextraction | 2026-09-09 |
Runs marked unenforced were measured on a suite that requires constrained decoding, on a server shown not to apply the schema: mlx_lm accepts it and ignores it, and a few other servers and models break it. Those numbers do not measure what the suite is designed to measure. They are kept and labelled rather than deleted.
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 | granite-guardian-4.1-8b-mlx-q8/ |
| Format | MLX 8-bit, 8.3 GB by du |
| Where from | made here 2026-09-09 ~12:10Z |
| Status | deleted 2026-09-09 14:08Z |
How it was made
- Input: the safetensors in the HF cache (Granite Guardian 4.1 8B BF16 safetensors (IBM, HF cache)), "source already cached".
- Command:
mlx-venv/bin/python -m mlx_lm convert --hf-path ibm-granite/granite-guardian-4.1-8b \
--mlx-path models/granite-guardian-4.1-8b-mlx-q8 -q --q-bits 8Group size default (not on the line).
- Tool: mlx_lm in
mlx-venv(run log: 0.31.3, recorded 2026-10-21 from the venv, not at run time). - Date: 2026-09-09 ~12:10Z.
- Size: 8.3 GB by du. Exact bytes not recorded.
- Why: to test the mixed_3_6 build against the same model at a near-lossless bit-width, with one variable changed. If the perplexities separated, the logit path was damaged (the recipe from a lab note). The assistant made it in answer to Alan's "can we identify why things are breaking". Alan's "is it the 3 6 that broke it or the plain mlx conversion" was answered by this build's screening run.
How it was served
| label | dates (UTC) | binary and version | flags | harness profile | why this label exists |
|---|---|---|---|---|---|
granite-guardian-4.1-8b-mlx-q8 | 2026-09-09 12:41Z–12:44Z | mlx-venv/bin/python -m mlx_lm server --model <path> | defaults | extraction | the quant-isolation cell. Command passed --label guardian-q8-mlxlm; results were filed under the path-derived name, and the run-log label is a later rename (a run-log listing on 2026-09-17 shows granite-guardian-4.1-8b-mlx-q8) |
No seat.
Decisions
- Made as a diagnostic. It showed the MLX conversion costs ~1.7 points at most.
- Deleted 2026-09-09 14:08Z on Alan's "ok remove the granite extras and the superseded models"; the assistant had listed it as "diagnostic only".
What went wrong
- Label not honoured (see Served as). No partial or void rows.
What happened to it
Converted 2026-09-09 ~12:10Z → benched 12:44Z → deleted 2026-09-09 14:08Z. No RESTORE entry. Re-creatable from the HF-cache safetensors with the command above.