UNCENSORED AI MODEL NEWS
Find your next model.
Uncensored and abliterated AI model news, cloud GPU costs and self-hosting research. Find a model worth running, then get help deploying it.
Release notes, source links, license details and reported benchmarks. Read the research. Choose your stack. Make it yours.
dealignai/Qwen3.8-Flash-Next-ABLITERATED-FP8
Qwen released Qwen3.8-Flash-Next on 24 August as the experimental preview of the Qwen4 architecture — 180B total parameters, ~6B active, hybrid Gated DeltaNet + micro-block sparse attention + 512-expert MoE + a 51B-parameter n-gram lookup table. Within 48 hours at least five editors raced to abliterate it. dealignai's FP8 build is the served, measured one: 100%/99.6% HarmBench-320 real-harm compliance at greedy decoding, MMLU 86.36→83.86, ~81% MTP draft acceptance. Jiunsong's BF16 edit is the architecture-aware method: 36 storage tensors → 6,168 logical output projections, 842-pair corpus. The catch: released vLLM/SGLang still don't support qwen4_exp.
Read the article → Model research ·dealignai/GLM-5.3-Flash-ABLITERATED-FP8
Zhipu's first natively multimodal GLM-5 (320B total / 18B active, hybrid KDA linear + sparse attention, MIT) landed 25 August. dealignai's CRACK edit removes refusals at the weight level in official FP8: 320/320 HarmBench-320 greedy compliance, 0 refusals, 30/30 at temp 1.0/top_p 0.95, MMLU 86.74→86.26 (−0.48pp), 163→211 tok/s with the cracked MTP head (75.9% acceptance) on 4×H200. NVFP4 twin (~165B) and an UNCENSORED-FP8 mirror with identical weights.
Read the article → Model research ·HangGlidersRule/Darkstar-Nemotron-3.5-Lightning-30B-A3B-Abliterated-BF16
NVIDIA's newest open model is a hybrid — Mamba-2 state-space blocks interleaved with MoE and sparse attention, 30B total / ~3B active, built for agent execution. Two weeks after launch, Darkstar removed the refusal direction from 3,126 residual-writing tensors with a numbered edit contract: max residual leakage 0.000160, 200/200 harmful compliance / 0/83 safe over-refusals publisher-measured, plus a ~22 GB NVFP4 twin that keeps Mamba tensors in BF16 and scores GPQA 71.2% on one GPU.
Read the article → Model research ·0xKitkat/Ornith-1.5-35B-A3B-Uncensored
DeepReinforce's self-improving Ornith-1.5-35B-A3B, uncensored three ways in 48 hours: a streamed task-vector transplant that adds Qwen3.6's measured uncensoring delta to Ornith's weights (102 tensors changed, vision and MTP intact, 0/16 heuristic refusals on Q4_K_M), plus two classic orthogonalization edits.
Read the article → Model research ·orcarouter/Qwen3.8-27B-Uncensored-FP8
The most-liked Qwen3.8 uncensored on Hugging Face (553 likes): orcarouter's lossless aggressive edit, block-FP8 with the vision tower at full precision, served gated on OrcaRouter at $0.40/$4.21 per 1M tokens.
Read the article → Model research ·insraq/Qwen3.5-4B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated
A source-linked field note on insraq's Heretic v1.4.0 decensor of EmperoAI's Qwen3.8-4B distill: what the 4.5B class is for, and what it trades away.
Read the article → Model research ·SHS-Lab/Muse-Glimmer-30B-Abliterated-Aggressive
A relaxed-KL LoRA de-abliteration of Meta's 29.8B agentic model claiming 0/100 refusals at ~1.7x KL drift — a mirror upload with the benchmarks left unmeasured.
Read the article → Model research ·Resggg/Kimi-K3-Abliterated-modal
An abliterated re-upload of Moonshot's 2.78T-parameter Kimi K3 — any-to-any, MXFP4, 96 shards, ~1.56 TB — with zero downloads and no deployment: the scale math behind why nobody can run it.
Read the article → Model research ·Goodoldjam/DiffusionGemma-26B-E38-Abliterated-NVFP4
Abliterating a model that doesn't generate autoregressively, then quantizing it to NVFP4: 0/402 target refusals and 1,053.64 tok/s aggregate on one RTX PRO 6000, 51.68 GB cut to 18.86 GB.
Read the article →Model research, dated at publication. Model licenses, publisher benchmarks and hosting estimates are specific to each article, not a live availability or price list. Reported zero-refusal results are test-specific, not a universal guarantee.