# Uncensored AI models & self-hosting guides

Uncensored and abliterated AI model news, cloud GPU costs and self-hosting research. Find a model worth running, then get help deploying it.

- [The base of the wave: Muse-Glimmer-30B's measured de-refusal](https://abliterated.cloud/blog/muse-glimmer-30b-abliterated/): 2026-08-12. jorkle's KL-measured LoRA-SFT de-refusal of Meta's dense 29.8B agentic model — the Normal twin of a three-repo family with its own GGUF ladder and base-reference quants.
- [A pentesting model, with the refusals taken out](https://abliterated.cloud/blog/huihui-cyberstrike-offsec-35b-abliterated/): 2026-08-10. An abliterated fine-tune of the CyberStrike-OffSec-35B offensive-security model: a Qwen3.6-35B-A3B tool-calling base, refusal weights edited out, no post-edit evaluations published.
- [Qwythos 9B: a model with three lives](https://abliterated.cloud/blog/qwythos-9b-claude-mythos-5-1m-abliterated/): 2026-07-18. Qwen architecture, Empero post-training and huihui-ai's final refusal-reduction pass, with a 1M label that deserves a closer look.
- [Inside Huihui-Qwen3.6: 256 experts and one refusal direction](https://abliterated.cloud/blog/qwen3-6-35b-a3b-abliterated/): 2026-07-18. The multimodal proof of concept: how huihui-ai applied abliteration to a 36-billion-parameter mixture-of-experts checkpoint, and why uncensored still needs caveats.
- [Ornith 397B: surgery on a model too large to hold at once](https://abliterated.cloud/blog/ornith-1-0-397b-abliterated-w4a16/): 2026-07-18. A shard-by-shard abliteration and W4A16 conversion of a 396.8-billion-parameter sparse model that still produces nearly 196 GiB of weights.
- [Ornith 35B: can self-scaffolding survive abliteration?](https://abliterated.cloud/blog/ornith-1-0-35b-abliterated/): 2026-07-18. DeepReinforce's coding scaffold, YuYu1015's corrected weights and the boundary between upstream and derivative benchmarks.
- [The workhorse: Huihui-Qwen3.6-27B-abliterated, four months in](https://abliterated.cloud/blog/huihui-qwen3-6-27b-abliterated/): 2026-04-23. 18,760 downloads in four months: what the community actually runs a dense 27B Qwen abliteration for, from red-teaming to quantization, and what its users report back.
- [The quiet classic: how Huihui-Qwen3.5-9B-abliterated became the small-model default](https://abliterated.cloud/blog/huihui-qwen3-5-9b-abliterated/): 2026-03-09. Abliterated Qwen3.5-9B (9,653,104,368 params, Apache-2.0) published 9 March 2026: 9,195 downloads and 125 likes on the base, with 58 downstream repos — GGUF/AWQ/MLX conversions and a preference-tuned Grimoire family — holding 64,688 combined downloads.
- [The MIT vision sleeper that resurfaced in August](https://abliterated.cloud/blog/huihui-glm-4-6v-flash-abliterated/): 2025-12-09. MIT-licensed text-side abliteration of Zhipu's GLM-4.6V-Flash vision model (10,292,777,472 params), published 9 December 2025, dormant for eight months, freshly re-quantized with vision-projector files on 17 August 2026.

[Get self-hosting help on Signal](https://signal.me/#p/+13103408213) · [All articles](https://abliterated.cloud/blog/)

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.