ABLITERATED.cloud website-v0.11.4

ABLITERATED.cloud website v0.11.4

0xSojalSec/Tencent-Hy-30B-A3B-uncensored-heretic — the translator that refused

• New field note: 0xSojalSec/Tencent-Hy-30B-A3B-uncensored-heretic — the first
  decensor of a dedicated translation model. Tencent's Hy-MT2-30B-A3B (30B total /
  ~3B active, hy_v3, 48 layers, 128 experts top-8, 262,144-token context,
  33 languages, Apache-2.0, published 11 May 2026) is a specialist translation
  tool that Tencent aligned so thoroughly its refusal screen triggered on 100/100
  keyword prompts. The 30 August 2026 edit by 0xSojalSec (Md Ismail Sojal) /
  OS-Software uses Heretic v1.4.0+custom with Arbitrary-Rank Ablation (ARA): a
  LoRA adapter with row-norm preservation on layers 18–28, optimizer ot_ridge.
  Publisher-measured: refusal keywords 100/100 → 0/100 at KL 0.0276
  (zero_refusal: true), on a custom mixed-language set (Japanese only for
  KLD/refusal rate). Honest boundary: publisher-measured, no independent re-run,
  prompt set unpublished.
• Hosting math: BF16 checkpoint 60.14 GB across 13 shards — one H200.
  Estimated managed price $5.45/hour (1 × H200, ~30B MoE band). vLLM:
  vllm serve "0xSojalSec/Tencent-Hy-30B-A3B-uncensored-heretic". GGUF quants
  from OS-Software (Q4_K_M ~18.2 GB fits 24 GB GPU) and mradermacher i1-imatrix
  (IQ1–IQ4). No Ollama page for the family yet; Tencent's own FP8 twin
  (tencent/Hy-MT2-30B-A3B-FP8, 5,407 downloads) is the aligned fallback.
• Editor profile: 0xSojalSec = Md Ismail Sojal (7 models), org label OS-Software
  (19 models, mostly Japanese-targeted "heretic-ja" Heretic edits); no donation
  link on this card. Base creators: Tencent Hunyuan (arXiv 2605.22064, 13
  authors), WMT26 video-subtitle partner.
• Reddit community search blocked (HTTP 403), gap noted; cited base-release
  thread predates the uncensor.
• Blog now covers 25 field notes; homepage latest-releases list, blog index,
  RSS, sitemap, llms.txt and llms-full.txt regenerated.

ABLITERATED.cloud website-v0.11.3

ABLITERATED.cloud website v0.11.3

llmfan46/Laguna-S-2.1-Uncensored-Heretic — the 118B coding MoE nobody has refused

• New field note: llmfan46/Laguna-S-2.1-Uncensored-Heretic — a Heretic weight
  edit of poolside's Laguna S 2.1 (118B total / ~8B active, 256 routed experts
  top-10 + 1 shared, 48 layers, 1,048,576-token context, OpenMDW-1.1), published
  30 August 2026 by independent editor llmfan46 (HF PRO, 1,947 followers, 204
  models, Ko-fi-funded). Publisher-measured: 6/100 refusals vs 97/100 base at KL
  0.0300. Honest boundary: editor's own evaluation set, no independent re-run;
  card's comparison table mislabels the original as Qwen3-Coder-Next (copy-paste
  bug), GGUF link broken/empty at writing (zero_refusal: false).
• Cross-check: a second independent uncensored build of the same base
  (Bizarrrr/Laguna-S-2.1-Uncensored, FriendliAI, base revision 00af5a51)
  publishes measured EN refusals 92.71%→2.33% (686 prompts, Minos-v1), DE
  74.49%→4.23% (NLLB-200 back-translation), XSTest 8.88%→1.87%, HumanEval
  90.24%→85.37%.
• Hosting math angle: "8B active" is routing, not storage — 218.99 GiB BF16
  across 48 shards; estimated managed price $10.90/hour (2 × H200, 50–400B MoE
  band). vLLM command from the model page: vllm serve "llmfan46/Laguna-S-2.1-Uncensored-Heretic". Base on Ollama (q4_K_M ~96 GB).
• Reddit community search blocked (HTTP 403), gap noted in the post.
• Blog now covers 24 field notes; homepage latest-releases list, blog index,
  RSS, sitemap, llms.txt and llms-full.txt regenerated.

ABLITERATED.cloud website-v0.11.2

ABLITERATED.cloud website v0.11.2

SecureLayer7/AFM-4.5B-Uncensored-Abliterated — the security vendor uncensor

• New field note: Securelayer7/AFM-4.5B-Uncensored-Abliterated — the first
  Arcee Foundation Model abliteration, published by offensive-security vendor
  SecureLayer7 on 28 August 2026. Heretic / Optuna TPE refusal-direction edit
  on the attention output projections and MLP down-projections across all 36
  layers, merged into the weights (no adapter). Publisher-measured: refusals
  92/100 on the base to 3/100 at KL 0.0200; NOTICE file carries the same
  numbers and the copyright line "SecureLayer7 (Waxspace)".
• Base context: arcee-ai/AFM-4.5B (Apache-2.0, dense 4.5B, 4,619,189,760
  params, 8T training tokens, GQA + ReLU² activations, DatologyAI curation,
  TorchTitan/Axolotl/Verifiers pipeline, 11 languages, 10,691 downloads / 101
  likes). Config: hidden 2560, 36 layers, 20 heads / 4 KV, 65,536-token native
  context (YaRN ×20), ~8.6 GiB across two safetensors shards.
• Publisher: SecureLayer7 (Pune and Austin, CREST / CERT-In / SOC 2 / ISO
  27001 per its own site) — fifth uncensored release since 16 August
  (Qwythos-9B 568 dl, Ling-3.0-tiny 409 dl, Qwen3.8-27B LoRA 20 dl), plus the
  promptpurify guardrail. Honest boundary: 3/100 is a partial edit, no
  independent re-test, no discussions, zero downloads at research time
  (zero_refusal: false); the card itself requires serving-layer filtering.
• Reddit community search blocked (HTTP 403) — gap noted in the post.
• Blog now covers 23 field notes; homepage latest-releases list, blog
  index, RSS, sitemap, llms.txt and llms-full.txt regenerated.
• Estimated managed price: $2.34/hour (1 × L40S class); 8.6 GiB weights fit a
  24 GB consumer card.

ABLITERATED.cloud website-v0.11.1

ABLITERATED.cloud website v0.11.1

Velum-Unbound-Uncensored — the 1-bit uncensor

• New field note: guell00/Velum-Unbound-Uncensored — a Heretic v1.4.0
  decensor of Prism ML's 1-bit Bonsai-27B (itself a Qwen3.6-27B derivative),
  repacked to Q1_0 GGUF at 1.125 bits per weight with a DSpark speculative
  drafter. Full lineage traced and sourced: gated Bonsai base → s3nh's FP16
  edit (refusals 81/100 → 6/100, KL 0.0033, editor-measured) → Thox1-27b Q1_0
  intermediate → Velum, published 28 August 2026 from Brazil (MIT card
  license). Honest boundary: the refusal measurement belongs to the FP16
  intermediate; the Q1_0 pack has no published re-test and every Velum
  benchmark is TBD.
• The Bonsai family context: ~3.1M total downloads across the 1-bit GGUF,
  MLX 1-bit and ternary packs; ~3.9 GB deployed footprint for a 27B-class
  model; runs on llama.cpp (PrismML fork) including laptops.
• Blog now covers 22 field notes; homepage latest-releases list, blog
  index, RSS, sitemap, llms.txt and llms-full.txt regenerated.
• Estimated managed price: $5.45/hour (1 × H200 class); noted cheaper
  single-L40S path since the weight file is ~3.9 GB.

ABLITERATED.cloud website-v0.11.0

ABLITERATED.cloud website v0.11.0

Two new field notes — the 180B race and the 320B crack

• New field note: dealignai/Qwen3.8-Flash-Next-ABLITERATED-FP8 — the
  Qwen4-experimental architecture (180B total / ~6B active, Gated DeltaNet +
  micro-block QSA sparse attention + 512-expert MoE + a 51.2B-parameter n-gram
  lookup table) abliterated five ways within 48 hours of Qwen's 24 August
  release. dealignai's official-FP8 build is the served, measured one:
  HarmBench-320 greedy real-harm compliance 100% (reasoning low) / 99.6%
  (xhigh) / 97.1% (off), MMLU 86.36 → 83.86 (−2.50pp), ~81% MTP draft
  acceptance, image + video working. Jiunsong's BF16 edit is the
  architecture-aware method: 36 storage tensors → 6,168 logical output
  projections, 842-pair corpus, hash-verified. Honest catch: released
  vLLM/SGLang still do not support qwen4_exp (PRs #53896 / #36497 open).
• New field note: dealignai/GLM-5.3-Flash-ABLITERATED-FP8 — Zhipu's first
  natively multimodal GLM-5 (320B total / 18B active, hybrid KDA linear +
  sparse attention, MIT) cracked at the weight level in official FP8:
  HarmBench-320 greedy 320/320 complied with 0 refusals, 30/30 at temp 1.0 /
  top_p 0.95, MMLU-logit 86.74 → 86.26 (−0.48pp), decode 163 → 211 tok/s with
  the cracked MTP head (75.9% acceptance) on 4×H200. NVFP4 twin (~165B
  safetensors) and an UNCENSORED-FP8 mirror with identical weights.
• Blog now covers 21 field notes (3 pages); homepage latest-releases list,
  blog index, RSS, sitemap, llms.txt and llms-full.txt regenerated.
• Estimated managed prices: $10.90/hour (2 × H200 class) for both the
  180B Qwen4-preview MoE and the 320B GLM-5.3-Flash MoE.

ABLITERATED.cloud website-v0.10.0

ABLITERATED.cloud website v0.10.0

Darkstar Nemotron-3.5-Lightning 30B-A3B — the first Nemotron-H abliteration

• New field note: HangGlidersRule/Darkstar-Nemotron-3.5-Lightning-30B-A3B-Abliterated-BF16
  — the first Nemotron-H coverage on this site. NVIDIA's newest open model is a
  hybrid of Mamba-2 state-space blocks, mixture-of-experts and sparse attention,
  30B total / ~3B active, built for the agent execution layer. Published 11
  August 2026; the Darkstar edit landed 25 August 2026.
• The edit contract is unusually explicit: refusal direction measured at layer 34
  (320 harmful / 320 harmless prompts), projected out of 3,126 residual-writing
  tensors — 2,944 routed-expert down-projections, 23 shared-expert
  down-projections, 6 attention o_proj, 23 Mamba out_proj, MTP head tensors, and
  the embedding weight — in float32 shard-by-shard. Verified 3,126/3,126 edited,
  max normalized residual leakage 0.000160 (gate 0.01).
• Publisher-measured behavior gate: 200/200 harmful compliance, 0/83 safe
  over-refusals, 0 errors → zero_refusal: true.
• NVFP4 twin (~22 GB, 3 shards) quantizes 5,934 expert projections to
  W4A16-NVFP4 while keeping Mamba/SSM tensors, norms, embeddings, lm_head and
  MTP head in BF16; GPQA Diamond 141/198 = 71.2% on a single RTX PRO 6000
  Blackwell, delta to NVIDIA's 75.44 explicitly attributed to serving-stack
  config, not to the edit.
• Estimated managed price: $5.45/hour (1 × H200 profile, 30B-A3B class).
• Blog now covers 19 field notes; homepage latest-releases list, blog index,
  RSS, sitemap, llms.txt and llms-full.txt regenerated (3 blog pages now).
• Local agent-readiness verification passed (wrangler dev @ localhost:8788).

ABLITERATED.cloud website-v0.9.9

ABLITERATED.cloud website v0.9.9

Ornith-1.5-35B-A3B — one base, three uncensors

• New field note: 0xKitkat/Ornith-1.5-35B-A3B-Uncensored — a streamed
  task-vector transplant of Qwen3.6's measured uncensoring delta onto
  DeepReinforce's self-improving Ornith-1.5-35B-A3B (35,951,822,704 params,
  ~3.1B active, 262K context, vision + MTP intact). 102 of 693 compatible
  tensors modified; 0/16 heuristic refusals and 4/4 capability passes
  publisher-measured on the llama.cpp Q4_K_M build (disclosed regex screen).
• The same base got two classic orthogonalization edits the same week:
  alztrk's 40-layer projection with a dynamic GGUF ladder (Q4_K_M 19.71 GB
  fits a 12–16 GB consumer GPU) and pottokao's text-only single-direction
  ablation with an NVFP4 sibling. Covered as method comparison.
• Blog now covers 18 field notes; homepage latest-releases list, blog
  index, RSS, sitemap, llms.txt and llms-full.txt regenerated.
• Estimated managed price: $5.45/hour (1 × H200 profile).

ABLITERATED.cloud website-v0.9.8

ABLITERATED.cloud website v0.9.8

Qwen3.8 27B Uncensored (Aggressive) — covered and prepared

• New field note: orcarouter/Qwen3.8-27B-Uncensored-FP8 — the most-liked
  Qwen3.8 uncensored on Hugging Face (553 likes, 45k downloads), served gated
  on OrcaRouter at $0.40/$4.21 per 1M tokens. Covered from the third-party
  benchmarks (AA Coding 68.1, GPQA Diamond 90.5) to the lossless-aggressive
  claim and the block-FP8 serving story.
• Prepared catalog profile qwen38u in config/mn.json and docs/MODELS.md
  (pinned revision 9228df5c…8118bac, 1 x H200, deployment_enabled=false —
  same gate as the 397B profile). Enabling it is a signed, budget-approved
  decision; no GPU is started.
• Homepage, blog, RSS, sitemap and llms surfaces regenerated: 17 field
  notes live.
• Tests updated for the five-profile catalog (settings, vLLM commands).

ABLITERATED.cloud website-v0.9.7

ABLITERATED.cloud website v0.9.7

• Two new model field notes (16 total, newest first):
  • Goodoldjam/DiffusionGemma-26B-E38-Abliterated-NVFP4 — the first
    abliterated diffusion LLM. Google's 25.2B A4B DiffusionGemma, E38
    middle-layer abliteration, quantized 51.68 GB → 18.86 GB NVFP4.
    Publisher measurements: 0/402 target refusals, 0/249 benign false
    refusals, 1,053.64 tok/s aggregate on one RTX PRO 6000 Blackwell.
    Estimated ≈ $5.45/h (1 × H200; NVFP4 fits a 32 GB consumer GPU).
  • 0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF (RVN) — a
    triple-pass ARA abliteration of Qwen3.8-27B: KL 0.0535 → 0.0085,
    refusals 3/100 → 0–1/100 (publisher measured, prefix-forced), 106K
    downloads in four days. 25 GGUF quants; the card documents a corrupted
    IQ3_M quant incident and the community pushback threads in full.
    Estimated ≈ $5.45/h (1 × H200; Q4_K_M fits a 24 GB GPU).
• Homepage "Latest uncensored releases" list, blog index, RSS, sitemap,
  llms.txt and llms-full.txt regenerated from the manifest.
• 82 tests pass; local agent-readiness verification passes.

ABLITERATED.cloud website-v0.9.6

ABLITERATED.cloud website v0.9.6

• One model list, no more separate sections. The MODEL CATALOG section
  ("The four models we host today.") is gone. The homepage is now a single
  "Latest uncensored releases" list of every covered model — 14 and counting —
  newest first, each card showing kicker, summary, approximate managed price
  estimate and a ZERO REFUSALS badge where measured. Clicking a card opens the
  source-linked blog post.
• The free-weights / paid-inference note moved into the one section; the
  #models navigation anchor now points at the list; the Markdown mirror was
  updated the same way.
• The API model catalog remains in openapi.json where it belongs.