Boring AI News

Qwen3-8B ZeroZero Heretic Removes Refusals and Slop

Published 8 November 2026

A modified, Apache 2.0-licensed Qwen3-8B model that reports 91% fewer refusals and 78% fewer slop words while retaining 13/15 fact and logic checks. Available in Safetensors, MLX, and GGUF formats, with no built-in safety filter.

models
Read the original on huggingface.co →

More from 8 November 2026

Every Monday we pick the five stories from the week worth keeping.

Read the latest AI Digest