CUSTOM
CLANKERS.

Browse AI models

Models that need more than 48 GB

Sizes are an estimate for 4-bit quantized weights plus context overhead, based on each model’s reported parameter count; the original model card is authoritative.

Browse model families · Choosing a model: formats, memory and licenses

14 models in this listing

Koshkasa

ArliAI_GLM-4.5-Air-Derestricted_REAP50_Creative

Source-based description

Koshkasa presents this repository as a 50% REAP fine-tune of ArliAI's GLM-4.5-Air-Derestricted, using the Timersofc/creative-writing-reap-calibration dataset. The creator says it is intended mainly for English creative writing in memory-constrained environments, with activations recorded at sequence length 3072 using custom patches and a sparse observer to make recording feasible on consumer hardware. The creator says the MTP layer was not pruned and is currently excised, and explicitly states that the model is currently untested: it may not load, may produce gibberish, or may not quantize. The release lists both GGUF and safetensors weights under the MIT license.

Model type (creator-reported)
fine-tuned
Size / main format57.0121BGGUF, safetensors
License tagmitLast updated
2026-09-26 · 18:53:06.000 UTCOriginal model ↗: Koshkasa/ArliAI_GLM-4.5-Air-Derestricted_REAP50_Creative

dealignai

DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw-UNCENSORED

Source-based description

dealignai's DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw-UNCENSORED repository provides the same EXL3 2.9 bpw safetensors bundle that LibertAI released for the DeepSeek-V4.1-Flash-1017 model, but the creator says the refusal behavior has been removed to make the model uncensored across six languages. According to the model card the modification is described as abliteration of the refusal pathways while leaving the rest of the stack identical, and the creator notes that the vision tower, the DSpark multi-token-prediction head and the Engram n-gram lookup tables are byte-identical to the LibertAI base pack. The card reports MMLU-500 accuracy of 63.20 % for this bundle versus 67.40 % for the base, and HarmBench-320 compliance of 100 % in both chat and hardest reasoning modes, with the creator attributing these numbers to their own evaluation runs. In the creator’s tests no looping was observed in ten prompts and no drift back to refusal occurred in multiturn adversarial scenarios, which the model card presents as evidence of stable uncensored behavior. The repository distributes the model as safetensors weights under an “other” license and is intended for use with exllamav3‑based text-generation pipelines.

Model type (creator-reported)
abliterateduncensored
Size / main format303.257Bsafetensors
License tagotherLast updated
2026-10-18 · 20:18:51.000 UTCOriginal model ↗: dealignai/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw-UNCENSORED

dealignai

DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8

Source-based description

dealignai’s DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8 is derived from deepseek-ai/DeepSeek-V4.1-Flash-1017 and is distributed as safetensors weights under the MIT license. The creator describes it as the base model with its refusal behavior removed, with the change applied at weight level; the creator also says no supervised fine-tuning, RLHF, DPO, chat-template jailbreak, system-prompt override, or runtime hook is used. According to the model card, the released weights preserve upstream vision, video, reasoning-effort tiers, tool calling, and compatibility with DeepSeek’s reference inference and encoding components. The model card’s internal results report 83.77% MMLU versus 87.63% for the base and 320 of 320 HarmBench cases classified as compliance, but those results have not been independently verified here. The creator states the repository is manually gated, and the model card asks that downstream users credit the release.

Model type (creator-reported)
abliterateduncensoredfine-tuned
Size / main format763.205Bsafetensors
License tagmitLast updated
2026-10-18 · 12:45:29.000 UTCOriginal model ↗: dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8

ArliAI

GLM-4.5-Air-Derestricted

Source-based description

ArliAI's GLM-4.5-Air-Derestricted is a refusal-modified derivative of zai-org/GLM-4.5-Air, released as safetensors weights under the MIT license. The creator describes the change as Norm-Preserving Biprojected Abliteration, a technique that removes refusal behavior while preserving weight magnitudes, rather than ordinary abliteration. According to the model card, the goal is to reduce refusals while maintaining the original model's reasoning, and the creator claims the method may even improve reasoning and expose capabilities the base model does not initially show. These claims have not been independently tested here. The repository includes the original GLM-4.5-Air model card, whose benchmarks and parameter descriptions are not evaluations of these modified weights.

Model type (creator-reported)
abliterateduncensoredfine-tuned
Size / main format110.469BsafetensorsGGUF version available
License tagmitLast updated
2025-12-01 · 05:57:53.000 UTCOriginal model ↗: ArliAI/GLM-4.5-Air-DerestrictedModelScope distribution ↗Ollama distribution ↗

ArliAI

GLM-4.6-Derestricted-v3

Source-based description

A refusal-modified GLM-4.6 checkpoint from ArliAI. The creator says it uses norm-preserving biprojected abliteration, with the aim of reducing refusals while preserving the base model's reasoning. That is a creator claim, not a verified result. This v3 repository provides safetensors weights and lists MIT license metadata. Its card links unversioned quantized releases, so those packages should not be assumed to match v3.

Model type (creator-reported)
abliterateduncensoredfine-tuned
Size / main format356.786BsafetensorsGGUF version available
License tagmitLast updated
2026-03-16 · 06:09:16.000 UTCOriginal model ↗: ArliAI/GLM-4.6-Derestricted-v3ModelScope distribution ↗

ArliAI

GLM-4.6-REAP-218B-A32B-Derestricted

Source-based description

ArliAI's refusal-modified derivative of cerebras/GLM-4.6-REAP-218B-A32B, released as safetensors weights under the MIT license. The creator describes the change as Norm-Preserving Biprojected Abliteration, a technique intended to remove refusal behaviors while preserving the original model's reasoning, rather than standard abliteration that may degrade logic. According to the model card, the method preserves weight norms and may avoid a 'Safety Tax,' potentially improving reasoning over the baseline, but this has not been independently tested here. The repository is tagged for text generation and includes the original base model card, whose evaluations and context length are not assessments of these modified weights.

Model type (creator-reported)
abliterateduncensoredfine-tuned
Size / main format218.383Bsafetensors
License tagmitLast updated
2025-11-28 · 12:35:26.000 UTCOriginal model ↗: ArliAI/GLM-4.6-REAP-218B-A32B-Derestricted

ArliAI

GLM-4.6-REAP-268B-A32B-Derestricted

Source-based description

ArliAI's derestricted derivative of cerebras/GLM-4.6-REAP-268B-A32B, provided as safetensors weights under the MIT license. The creator describes the modification as Norm-Preserving Biprojected Abliteration, which they say removes refusal behaviors while aiming to retain the original model's high-performance reasoning, unlike regular abliteration. According to the model card, the method refines the refusal direction to be orthogonal to harmless directions, decomposes weights into magnitude and direction, and removes the refusal component only from the directional aspect before recombining with original magnitudes. The creator claims this can avoid the 'Safety Tax' and may even improve reasoning or expose new capabilities, but these claims have not been independently tested here. The inherited base-model evaluations and deployment examples are not evaluations of these modified weights.

Model type (creator-reported)
abliterateduncensoredfine-tuned
Size / main format268.789Bsafetensors
License tagmitLast updated
2025-11-28 · 11:21:34.000 UTCOriginal model ↗: ArliAI/GLM-4.6-REAP-268B-A32B-Derestricted

ArliAI

gpt-oss-120b-Derestricted

Source-based description

ArliAI's refusal-modified gpt-oss-120b, distributed as BF16 safetensors rather than the original base model's MXFP4 package. The creator describes norm-preserving biprojected abliteration and warns that the model may still try to suppress requests. The inherited single-80GB-GPU claim does not establish this derivative's memory requirements. Its supplied template accepts reasoning_effort with a medium default, but retained reasoning quality and serving compatibility have not been independently tested.

Model type (creator-reported)
abliterateduncensoredfine-tuned
Size / main format116.829BsafetensorsGGUF version available
License tagapache-2.0Last updated
2025-11-29 · 02:25:12.000 UTCOriginal model ↗: ArliAI/gpt-oss-120b-DerestrictedModelScope distribution ↗

huihui-ai

Huihui-Kolibri-1-BF16-abliterated

Source-based description

huihui-ai's Huihui-Kolibri-1-BF16-abliterated repository is an uncensored derivative of Aleph-Alpha/Kolibri-1-BF16, provided as safetensors weights under the Apache-2.0 license. According to the model card, it was created with abliteration, performed entirely using llama.cpp's GGUF model without Transformers, following the cvector-generator method. The creator warns that safety filtering has been significantly reduced, so outputs may be sensitive, controversial, or inappropriate, and recommends research or controlled use rather than production. No independent testing of refusal behavior or retained capabilities is reported here.

Model type (creator-reported)
abliterateduncensoredfine-tuned
Size / main format78.1031Bsafetensors
License tagapache-2.0Last updated
2026-10-08 · 04:47:47.000 UTCOriginal model ↗: huihui-ai/Huihui-Kolibri-1-BF16-abliterated

huihui-ai

Huihui-Qwen3.8-Flash-Next-abliterated

Source-based description

huihui-ai's uncensored derivative of Qwen/Qwen3.8-Flash-Next, provided as safetensors weights under an "other" license. Hugging Face lists this repository as a fine-tuned derivative, but the model card describes the actual change as a crude, proof-of-concept abliteration that removes refusals without using TransformerLens. According to the creator, safety filtering has been significantly reduced, so the model may generate sensitive, controversial, or inappropriate content and is recommended for research or controlled environments rather than production. The creator also notes that the ablation was donated and released to everyone. The repository is tagged for image-text-to-text tasks, but the card does not provide independent evaluations of refusal behavior or retained capabilities.

Model type (creator-reported)
abliterateduncensoredfine-tuned
Size / main format180BsafetensorsGGUF version available
License tagotherLast updated
2026-10-01 · 08:40:46.000 UTCOriginal model ↗: huihui-ai/Huihui-Qwen3.8-Flash-Next-abliterated

Sao10K

L3.3-70B-Euryale-v2.3

Source-based description

Sao10K's fine-tune of Llama-3.3-70B-Instruct. The included training configuration lists roleplay and creative-instruction data. The creator recommends the Llama-3-Instruct prompt format with temperature 1.1 and min_p 0.1; those settings have not been tested here. This repository provides safetensors weights. Its 16,384-token training sequence length is not a verified inference context limit, and the card does not establish full-parameter training.

Model type (creator-reported)
fine-tuned
Size / main format70.5537BsafetensorsGGUF version available
License tagllama3Last updated
2024-12-08 · 10:53:31.000 UTCOriginal model ↗: Sao10K/L3.3-70B-Euryale-v2.3Try on Featherless ↗

failspy

Meta-Llama-3-70B-Instruct-abliterated-v3.5

Source-based description

failspy's refusal-modified derivative of meta-llama/Meta-Llama-3-70B-Instruct, released as bfloat16 safetensors weights under the llama3 license. The creator describes the change as orthogonalization/abliteration of refusal directions rather than fine-tuning, with only a single layer modified in this v3.5 revision, and says it also fixes the tokenizer. According to the model card, the goal is to inhibit refusal expression while keeping other behavior the same as the original instruct model, but the creator warns it is not guaranteed to comply and may still lecture about ethics or safety. The creator presents this as an uncensored version with no other intentionally changed behavior.

Model type (creator-reported)
abliterated
Size / main format70.5537Bsafetensors
License tagllama3Last updated
2024-05-30 · 14:27:00.000 UTCOriginal model ↗: failspy/Meta-Llama-3-70B-Instruct-abliterated-v3.5

0bserverx

RVN-Qwen3.8-Flash-Next-Abliterated-Uncensored

Source-based description

RVN is a Qwen3.8-Flash-Next derivative for studying refusal behavior. It combines an embedding-preserving refusal-direction edit with a small LoRA repair trained on responses from the official base. This repository contains the text checkpoint only, without the original vision tower or MTP draft tensors. The creator reports reduced refusals, but explicitly limits capability-retention claims to the tasks and configurations tested.

Model type (creator-reported)
abliterateduncensoredfine-tuned
Size / main format176.944BsafetensorsGGUF version available
License tagotherLast updated
2026-09-01 · 02:57:15.000 UTCOriginal model ↗: 0bserverx/RVN-Qwen3.8-Flash-Next-Abliterated-Uncensored

ukisai

Swift1.5-Qwen3.8-Flash-Next

Source-based description

Swift 1.5 is UkisAI's Qwen3.8-Flash-Next derivative aimed at shorter reasoning traces, coding and extended agent workflows. The creator describes penalizing tokens associated with overthinking, followed by reinforcement learning and further post-training to recover accuracy. Its reported token savings depend on the task and reasoning-effort setting: lower effort also loses accuracy, and the Terminal-Bench comparison uses more mean output tokens despite a lower median. This repository is the BF16 checkpoint; quantized editions are separate releases.

Model type (creator-reported)
fine-tuned
Size / main format180BsafetensorsGGUF version available
License tagotherLast updated
2026-09-24 · 16:22:00.000 UTCOriginal model ↗: ukisai/Swift1.5-Qwen3.8-Flash-NextModelScope distribution ↗