DefiledAI
WEEKLY DIGEST
New model drops, recent articles, tutorial updates, and backend news. Subscribe via RSS.
Week of May 31, 2026LATEST
New Model Drops4 this week
Llama 3.1 70B Abliterated v3ABLITERATED
by failspy
Updated quality retention on reasoning chains.
Qwen 2.5 72B Uncensored GGUFUNCENSORED
by bartowski
Full quant pack Q2_K through Q8_0.
Mistral Nemo 12B AbliteratedABLITERATED
by bartowski
128K context retained.
DeepSeek R1 8B AbliteratedABLITERATED
by Community
6GB VRAM minimum.
Recent Articles
Models
Llama 3.1 70B Uncensored
May 30, 2026
Leaderboard
Best Abliterated Models 2026: Community Rankings
2026-05-30
Research
Abliteration Explained: How Refusal Removal Works
2026-05-29
Model Analysis
Llama 3.1 70B: The Complete Local Inference Guide
2026-05-28
Quantization
Q4_K_M vs IQ3_M: Quantization Quality Analysis
2026-05-27
Hardware
Dual RTX 3090 NVLink: The 70B Inference Workstation
2026-05-26
Benchmarks
ExLlamaV2 vs llama.cpp: Which Backend Is Faster in 2026?
2026-05-25
Model Analysis
DeepSeek V3: Running a 671B MoE Model Locally
2026-05-24
New Tutorials
Backend & Tool Updates
›ExLlamaV2 v0.2.4 — ~12% throughput improvement on Q4_K_M for Ada GPUs
›llama.cpp adds experimental Q4_K_XS format — smaller than Q4_K_S at similar quality
›Community finds secondary refusal vectors in Llama 3.1 — improved abliteration script posted
›ROCm 6.2 released — closes AMD/CUDA inference gap on RX 7000 series