Weightless: tiny GLP steering vectors abliterate model refusals without touching weights
EAccelerate_42 · x · 2026-11-09
The weightless project showcases GLP — 'abliteration without the weights': a small steering vector subtracts a component at a projection site to steer refusal behavior while base weights stay untouched, enable/disable per request. It supports models from 2.6B to 2.9T params with 14 GLP vectors across 9 vendors. Featured project spark-ds41f deploys DeepSeek V4.1 Flash on a DGX Spark cluster via vLLM (95.9 tok/s generation, 6,370 tok/s prefill, 1.2s TTFT at 8K) with the GLP-27 vector (layers 11-37, α=3.5) serving natural and abliterated modes side by side. Ships with a stdlib-only TUI wizard for setup, steering, and deployment.
More from Models
- Claude Fable 5.5 needs a 70 on AA index to justify replacing Opus 5.5 — haider1 · 2026-11-09
- Leaked checkpoints suggest Google now ships a major Gemini release every month — lyraxana · 2026-11-09
- HF's Niels Rogge releases 'Getting Started with Local AI' slides: quantization, MoE, local agents — NielsRogge · 2026-11-09
- InstaDeep CEO: Anthropic's unreleased Models 3 and 4 sit halfway to AI2027's Agent-2 — connoraxiotes · 2026-11-09
- 27B open-weight decision model Drex Max takes #1 on Artificial Analysis Decision Index — dr_cintas · 2026-11-09
- JEV models get full support in Hugging Face transformers, now standardized — art_zucker · 2026-11-09