OPEN LINK ↗
// 1h agoMODEL RELEASE
Qwen3-4B Deslop Cuts AI Tics
Qwen3-4B Deslop is an open 4B fine-tune that removes common AI writing habits such as em dashes, rhetorical formulas, and triplet lists. Its creator reports reducing a style score from 121 to 12.5 per 1,000 words while preserving factual performance, with a 2.5 GB quantized build for local use.
// ANALYSIS
This is a compelling demonstration that model style can be edited cheaply, though the results remain heavily dependent on self-authored detectors and small evaluations.
- –LoRA distillation plus targeted neuron edits make the approach practical on a MacBook Air.
- –The model reportedly retains 30/30 factual quiz answers and restores task adherence to 80%.
- –The strongest gains target recognizable AI prose patterns, especially em dashes and “not X, but Y” constructions.
- –The voice can become flat, overly literal, and prone to invented names or places.
- –The experiment suggests a reusable recipe for removing corporate jargon or enforcing brand-specific writing styles.
// TAGS
qwen3-4b-deslopllmsmall-llmopen-weightsfine-tuningtraininglocal-first
DISCOVERED
1h ago
2026-10-19
PUBLISHED
5h ago
2026-10-19
RELEVANCE
9/ 10
AUTHOR
erikkannike