FreshContextA context lab · Fife
Drawing
02 · Benchmarks
Sheet
2 of 3
Rev
A · 2026-10-01

Granite Guardian 4.1 8B · file← Granite Guardian 4.1 8B

MLX bf16

An MLX copy with no rounding, as the control. Deleted the same day. Unquantised conversion, made here.

Size
16.0 GB
Status
Deleted 9 Sep 2026
Scored runs
1
Smallest Mac
32 GB estimate, 8K context
Results1 scored · 95% intervals
SuiteScore95% interval PassedRuntime · profileRecorded
screeninggranite-guardian-4.1-8b-mlx-bf1698.3%91.1–99.7Passed 59/60mlx_lmextraction2026-09-09

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.

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.

Filegranite-guardian-4.1-8b-mlx-bf16/
FormatMLX, no quantisation, 16 GB by du
Where frommade here 2026-09-09 ~12:49Z
Statusdeleted 2026-09-09 14:08Z

How it was made

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-bf16

(no -q: unquantised; the source dtype is bfloat16, so the label "fp16" in the script name is loose).

  • Tool: mlx_lm in mlx-venv (run log: 0.31.3, recorded later from the venv).
  • Date: 2026-09-09 ~12:49Z. Size: 16 GB by du; exact bytes not recorded.
  • Who asked: Alan, "can we do an fp16 -> mlx test, also omlx has native kernels for various model families, without them it defaults to generic implementation".

How it was served

labeldates (UTC)binary and versionflagsharness profilewhy this label exists
granite-guardian-4.1-8b-mlx-bf162026-09-09 ~12:50Z–12:54Zmlx-venv/bin/python -m mlx_lm server --model <path>defaultsextractionthe fp16 control. Command passed --label guardian-bf16-mlxlm; results were filed under the path-derived name, and the run-log label is a later rename

No seat.

Decisions

  • Made as the no-quantisation control; equal to 8-bit (59/60), so "no measurable conversion cost and no 8-bit cost".
  • Deleted 2026-09-09 14:08Z with the 8-bit.

What went wrong

  • Label not honoured (see Served as).

What happened to it

Converted 2026-09-09 ~12:49Z → benched → deleted 2026-09-09 14:08Z. No RESTORE entry. Re-creatable from the HF-cache safetensors.