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.
OS-Software/gemma-4-31B-it-qat-q4_0-uncensored-heretic-GGUF
Google's Gemma 4 31B (30.7B dense, 60 layers, 256K context, text and image input, Apache 2.0) has been uncensored repeatedly in 2026. coder3101's April Heretic v1.2.0 + ARA edit reports refusals 15/100 vs 99/100 at KL 0.0434; wnfldchen's 14 August rebuild on the Q4_0 quantisation-aware checkpoint reports 8/100 vs 99/100 at KL 0.0900 with a reproduction kit and a W4A16 capability table; OS-Software's 8 September release, made with a Heretic v2.0.0.dev0+custom development build, reports 0/100 against 100/100 for the stock checkpoint at KL 0.0083, shipping a 17.29 GB Q4_0 GGUF plus a 1.20 GB or 0.81 GB vision projector. EXL3 and Q4_0 requants, an expert-pruned 26B A4B variant, an E4B decensor (7/100 at KL 0.0043) and a separate Madras1 orthogonalization line complete the picture. All refusal and KL numbers are publisher-measured; no GPU was rented and no inference was run for this article.
Read the article → Model research ·dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8
DeepSeek released V4.1-Flash on 10 September 2026 (552B backbone, 8B/16B active per token, causal encoder-decoder, 384 routed experts, Engram memory, 1M-token context, MIT) and three independent uncensored builds followed within 19 hours. s-zaizen's Heretic edit (commit 3521f864) cut refusals 97/100 to 24/100 on the same 100-prompt keyword screen. msuiche's 800 KB GLP-39 control vector, gated, raised compliance 4/32 to 24/32 on a refusal set and 5/32 to 31/32 on a cyber set at alpha 0.5, with the card showing alpha 2.0 regresses and damages English. dealignai's CRACK-brand weight edit reports 320/320 HarmBench-320 compliance versus 137/320 stock at reasoning off, MMLU 86.96% to 82.74% (-4.22pp; -1.1pp excluding ethics). No released runtime serves deepseek_v41 yet: an SGLang preview branch and community forks do. Publisher-measured claims; no GPU was rented and no inference was run for this article.
Read the article → Model research ·insraq/MiniCPM5-2B-heretic-abliterated
OpenBMB shipped MiniCPM5-2B on 6 September 2026 (2,516,756,480 params, dense LlamaForCausalLM, 131,072-token context, Apache-2.0, open UltraData training sets) with a card claiming 2B-class SOTA at 53.9 average, above every listed 4B rival. Stock, insraq's screen measured 99 refusals of 100 prompts. Thirty-four hours later his Heretic v1.4.0 edit, trial 254 of an Optuna Pareto search, measured 5/100 at KL 0.0391 on the same screen, and it ships a reproduction kit: base commit 3497c460, pinned prompt datasets, RNG seed, full Optuna journal and SHA256SUMS, so a rebuild is byte-checkable. GGUF (F16 5.04 GB to IQ3_M 1.23 GB) and Apple MLX 4-bit packs followed within two days. Publisher-measured claims; no GPU was rented and no inference was run for this article.
Read the article → Model research ·soyaakinohara/Spark-X2.5-4B-Heretic
XHToken's SparkLLM shipped Spark-X2.5-4B (4,112,079,360 params) and -1.7B on 24 Aug 2026: Apache-2.0, hybrid attention (one full-attention layer per three sliding-window layers), up to 1M-token native context claimed, 200+ languages, trained on Huawei Ascend, and vendor agent benchmarks (τ³-bench 30.4 vs Qwen3.5-9B 9.3, AIME 2026 90.7) that outrank 9B/12B rivals. Stock, the small models refuse hard: 2,014/3,911 (51.5%) marker refusals on the 1.7B, 186/300 sealed, 100/100 sealed adversarial. Ten days after the drop the uncensors landed measured: soyaakinohara's BF16 Heretic of the 4B (58→3/100 refusals, KL 0.0118, base revision pinned, JP-adapted twin and a GGUF ladder to Q4_K_M 2.5 GiB) and darioooooo0o's 1.7B ablation (0 real refusals on 337 eye-audited generations vs 186/300 base, sealed sets, bucket ruler), plus a methodology argument about how to count refusals on thinking models. Runtime reality check: upstream llama.cpp lacks spark2_5; the vendor's own fork serves it, and an open GitHub issue asks for an official 1M-context reproduction config.
Read the article → Model research ·apetersson/DeepSeek-V4-Flash-Vision-Exp-Abliterated
DeepSeek's first open multimodal model (285B total / 13B active MoE, 1M context, 32-layer vision tower, DSpark self-draft) landed 31 Aug 2026 at 06:16 UTC; apetersson's rank-1 abliteration (33 attention-output tensors, layers 10-42, strength 3.5, L2-preserving) followed at 17:12 UTC. Text and image smoke tests pass, broader evaluation pending; ~202 GB FP8-native, ~$10.90/h managed on 2 x H200. Within a day s-zaizen grafted the edit onto NVFP4 (8/8 safe prompts, 0 refusals, GSM8K 99/100) and audreyt onto an IQ2_XXS ds4 GGUF; every derivative pins the donor revision.
Read the article → Model research ·0xSojalSec/Tencent-Hy-30B-A3B-uncensored-heretic
The first decensor of a dedicated translation model: Tencent's 33-language Hy-MT2-30B-A3B (30B total / ~3B active, 128 experts, hy_v3, 262K context, Apache-2.0) edited by 0xSojalSec/OS-Software with Heretic ARA (LoRA adapter, row-norm preservation, layers 18–28): refusal keywords 100/100 → 0/100 at KL 0.0276 (publisher measured), 60.14 GB BF16 across 13 shards, ~$5.45/h managed on 1×H200, GGUF quants down to Q4_K_M 18.2 GB.
Read the article → Model research ·llmfan46/Laguna-S-2.1-Uncensored-Heretic
Poolside's 118B-A8B agentic coding MoE (1M context, 256 routed experts, OpenMDW-1.1) uncensored by independent editor llmfan46 with Heretic: 6/100 refusals vs 97/100 at KL 0.0300 (publisher-measured), 219 GB BF16 across 48 shards, ~$10.90/h managed on 2×H200. A second independent build of the same base (Bizarrrr, on FriendliAI) measures EN refusals 92.71%→2.33%, DE 74.49%→4.23%, HumanEval 90.24%→85.37%.
Read the article → Model research ·Securelayer7/AFM-4.5B-Uncensored-Abliterated
The first Arcee Foundation Model abliteration, published by an offensive-security vendor: SecureLayer7's Heretic/Optuna TPE edit cuts AFM-4.5B refusals from 92/100 to 3/100 at KL 0.0200, merged into the weights. Apache-2.0, 4.5B dense, ~8.6 GiB, fits one L40S.
Read the article → Model research ·guell00/Velum-Unbound-Uncensored
The Bonsai family, the most downloaded 1-bit experiment of the summer (1.95M downloads on the MLX pack alone), gained an uncensored branch this morning: a Heretic v1.4.0 decensor of Bonsai-27B (6/100 refusals vs 81/100, KL 0.0033, editor-measured at FP16) repacked to Q1_0 at 1.125 bits per weight. Velum-Unbound-Uncensored is two files: a ~4.35 GiB Q1_0 GGUF plus a ~1.81 GiB dspark speculative drafter, MIT-licensed, from Brazilian independent publisher guell00. The honest catch: every benchmark on the card is TBD, the refusal number belongs to the FP16 intermediate, and nobody has re-measured the 1-bit pack. Runs in ~3.9 GB on llama.cpp (PrismML fork); an H200 is overkill, a laptop suffices.
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.