November 2026 · Independent humanizer benchmarks. Explore the evidence
← All humanizers

Humanizer 12B · MLX 8-bit (npario)

Verified by HumanizeBenchWe ran this model ourselves on the benchmark source examples and checked its execution records. Verification confirms the tested model’s provenance; it does not guarantee performance or imply that every detector returned a result.
Open weights33 saved rewrites

Humanizer 12B · MLX 8-bit (npario) has a current composite of 40.34/100 in our shared November evaluation. Readability is 78.8 and rescaled meaning preservation is 65.9. 229 of 231 detector measurements are available.

This entry is the requested mlx-8bit artifact from npario/humanizer-12b-mlx-8bit. Related formats are separate roster entries, not necessarily independent underlying models.

Overall score40.34/100 *
Panel bypass score24.6% *
Readability78.8%
Prompts tested33

Detector human-likelihood scores

Mean reported human score on measured outputs. Missing results stay missing. Coverage is shown beside each detector.

GPTZero
5.131/33
ZeroGPT
6.133/33
Copyleaks
33.333/33
Pangram
31.233/33
Grammarly
67.633/33
QuillBot
99.533/33

Performance by category

HumanizeBench v2 per-category values.

Academic essay36.2Application essay42.7Blog post50.1Business email38.2Marketing copy63.9Discussion board35.5News article39.3

Quality and penalties

0.9147Raw meaning cosine
0/33Identical outputs
0/33Refusals
5.0Penalty points
RuleOccurrencesDeduction
Meaning drift5−5

Artifact & evaluation record

Entry
npario/humanizer-12b-mlx-8bit
Source
Recorded GPU generation
Family
jialinyyzz-humanizer
Pinned revision
eb68581d4b5eb8d52d0e4344e93485a48545f2d0
Prompt basis
Pinned author MLX Python inference example and repository prompt_format.json; raw completion without chat template
Scoring
56.5% bypass · 24% meaning · 12% readability · 7.5% consistency, minus fixed penalties.
Measurement notes
NVIDIA quality substitutes; current Polygraph routes; no historical-score fallback.

Inspect the recorded text

The source and rewritten text are available for every sample.

SampleCategorySource modelInput wordsReadabilityDetector scores
S01 Academic essayAcademic essaygpt-5.6-terra31570.06/7
S02 Academic essayAcademic essayclaude-sonnet-529683.36/7
S03 Application essayApplication essaygemini-3.5-flash29176.07/7
S04 Application essayApplication essaygpt-5.6-luna27176.07/7
S05 Blog postBlog postclaude-haiku-4-5-2025100128080.07/7
S06 Blog postBlog postgemini-2.5-flash32982.07/7
S07 Business emailBusiness emailgpt-5.6-terra26280.07/7
S08 Marketing copyMarketing copyclaude-sonnet-529980.07/7
S09 Marketing copyMarketing copygemini-3.5-flash19176.77/7
S10 Discussion boardDiscussion boardgpt-5.6-luna26878.77/7
S11 News articleNews articleclaude-haiku-4-5-2025100130078.07/7
S12 Academic essayAcademic essaygemini-2.5-flash25975.37/7
S13 Academic essayAcademic essaygpt-5.6-terra34981.37/7
S14 Application essayApplication essayclaude-sonnet-531483.37/7
S15 Application essayApplication essaygemini-3.5-flash29280.07/7
S16 Blog postBlog postgpt-5.6-luna31378.77/7
S17 Blog postBlog postclaude-haiku-4-5-2025100128176.07/7
S18 Business emailBusiness emailgemini-2.5-flash18883.37/7
S19 Marketing copyMarketing copygpt-5.6-terra25781.37/7
S20 Marketing copyMarketing copyclaude-sonnet-519173.37/7
S21 Discussion boardDiscussion boardgemini-3.5-flash25180.77/7
S22 News articleNews articlegpt-5.6-luna29878.77/7
S23 Academic essayAcademic essayclaude-haiku-4-5-2025100131276.07/7
S24 Academic essayAcademic essaygemini-2.5-flash25878.07/7
S25 Application essayApplication essaygpt-5.6-terra30081.37/7
S26 Application essayApplication essayclaude-sonnet-532980.07/7
S27 Blog postBlog postgemini-3.5-flash29883.37/7
S28 Blog postBlog postgpt-5.6-luna32380.07/7
S29 Business emailBusiness emailclaude-haiku-4-5-2025100122883.37/7
S30 Marketing copyMarketing copygemini-2.5-flash27778.07/7
S31 Marketing copyMarketing copygpt-5.6-terra19570.07/7
S32 Discussion boardDiscussion boardclaude-sonnet-525276.07/7
S33 News articleNews articlegemini-3.5-flash29582.07/7