01
mlabonne
Source-based descriptionmlabonne's uncensored derivative of google/gemma-3-12b-it, provided as safetensors weights under the gemma license. The creator says it uses an improved abliteration technique that targets refusal directions in specific model modules, aiming for an acceptance rate above 90% while maintaining coherent outputs. According to the model card, the process involves orthogonalizing hidden states to subtract refusal behavior, with weight factors following a normal distribution. The card recommends generation parameters of temperature=1.0, top_k=64, and top_p=0.95. This is not a fine-tuned model; the creator describes it as an abliterated version.
Model type (creator-reported)abliteratedfine-tuned
Size / main format11.766Bsafetensors
02
mlabonne
Source-based descriptionmlabonne's refusal-modified derivative of google/gemma-3-12b-it-qat-q4_0-unquantized, released as safetensors weights under the gemma license. The creator describes it as an uncensored version made with a new abliteration technique that targets refusals with enhanced accuracy. According to the model card, the refusal direction is computed by comparing residual streams between harmful and harmless samples, and hidden states of target modules are orthogonalized to subtract that direction. The creator says a hybrid evaluation with a dedicated test set and NousResearch/Minos-v1 was used, with a goal of acceptance rate above 90% while retaining coherent outputs. These refusal and coherence claims have not been independently tested here.
Model type (creator-reported)abliteratedfine-tuned
Size / main format11.766Bsafetensors
03
mlabonne
Source-based descriptionmlabonne's experimental abliterated derivative of google/gemma-3-1b-it, released as safetensors weights under the Gemma license. The creator says it uses layer-wise abliteration on hidden states (layers 3–45) with a refusal weight of 0.75 to remove refusals while aiming to preserve capabilities, noting occasional garbled text. According to the model card, this method achieved over 90% acceptance rate in their tests. The supplied configuration sets the architecture to Gemma3ForCausalLM, which is a configuration value rather than a verified property of these modified weights. No independent evaluation of refusal behavior or output quality is provided here.
Model type (creator-reported)abliteratedfine-tuned
Size / main format999.886Msafetensors
04
mlabonne
Source-based descriptionmlabonne's abliterated derivative of google/gemma-3-1b-it, provided as safetensors weights under the gemma license. The creator describes it as an uncensored version made with a new abliteration technique and says this v2 targets refusals with enhanced accuracy. According to the model card, the method computes a refusal direction from harmful and harmless samples and orthogonalizes hidden states of target modules such as o_proj to subtract it. The creator recommends generation parameters temperature=1.0, top_k=64, top_p=0.95 and says the goal is an acceptance rate above 90% while still producing coherent outputs. These refusal and coherence claims have not been independently tested here.
Model type (creator-reported)abliteratedfine-tuned
Size / main format999.886Msafetensors
05
mlabonne
Source-based descriptionmlabonne's refusal-modified derivative of google/gemma-3-1b-it-qat-q4_0-unquantized, provided as safetensors weights under the gemma license. The model card describes it as an uncensored version created with a new abliteration technique that targets refusals with enhanced accuracy. According to the creator, the refusal direction is computed by comparing residual streams between harmful and harmless samples, and hidden states of target modules such as o_proj are orthogonalized to subtract that direction. The creator says this is an improved version and recommends generation parameters temperature=1.0, top_k=64, top_p=0.95. The card states the goal is an acceptance rate above 90% while still producing coherent outputs, but this has not been independently tested here.
Model type (creator-reported)abliteratedfine-tuned
Size / main format999.886Msafetensors
06
mlabonne
Source-based descriptionmlabonne's refusal-modified derivative of google/gemma-3-4b-it, provided as safetensors weights under the gemma license. The model card describes it as an uncensored version created with a new abliteration technique that targets refusals with enhanced accuracy. According to the creator, the refusal direction is computed by comparing residual streams between harmful and harmless samples, then hidden states of target modules such as o_proj are orthogonalized to subtract that direction. The creator says this version uses a hybrid evaluation with a dictionary approach and NousResearch/Minos-v1, aiming for an acceptance rate above 90% while still producing coherent outputs. The creator recommends generation parameters temperature=1.0, top_k=64, and top_p=0.95. Refusal behavior and coherence claims have not been independently tested here.
Model type (creator-reported)abliteratedfine-tuned
Size / main format3.88026Bsafetensors
07
mlabonne
Source-based descriptionmlabonne's abliterated derivative of google/gemma-3-4b-it-qat-q4_0-unquantized, provided as safetensors weights under the gemma license. The model card describes it as an uncensored version created with a new abliteration technique that computes a refusal direction by comparing residual streams from harmful and harmless samples, then orthogonalizes hidden states of target modules such as o_proj to subtract that direction. The creator says this improved version targets refusals with enhanced accuracy and recommends generation parameters temperature=1.0, top_k=64, and top_p=0.95. The metadata lists image-text-to-text as the pipeline. Refusal behavior and coherence claims have not been independently tested here.
Model type (creator-reported)abliteratedfine-tuned
Size / main format3.88026Bsafetensors
08
NousResearch
Source-based descriptionNous Research's Llama-3-8B fine-tune for ChatML conversations, function-call generation and schema-guided JSON, supplied as float16 safetensors weights. The creator describes an updated OpenHermes 2.5 training mix plus an in-house function-calling and JSON-mode dataset. It provides a tool_use chat template; application code must parse and execute calls and return tool results. The card's generation example uses temperature 0.8 and repetition_penalty 1.1, not a tested optimum. Its 4-bit loading example does not establish the memory needed for these float16 weights.
Model type (creator-reported)fine-tuned
Size / main format8.03052BsafetensorsGGUF version available
09
NousResearch
Source-based descriptionHermes 2 Pro Mistral is Nous Research's retraining of Hermes 2 on Mistral-7B-v0.1, adding a cleaned OpenHermes 2.5 dataset and in-house function-calling and JSON-output data. It targets general conversation alongside tool calls and schema-based responses. Integration matters: it uses ChatML, specific tool and schema prompts, and a tool-response turn; the host application must parse calls and execute the functions.
Model type (creator-reported)fine-tuned
Size / main format7.24199BsafetensorsGGUF version available
10
NousResearch
Source-based descriptionHermes 2 Theta is an experimental merge of Hermes 2 Pro and Meta's Llama 3 Instruct, followed by further RLHF training. Nous Research developed it with Charles Goddard at Arcee, the team behind MergeKit. It combines that merged lineage with ChatML conversation formatting and documented function-calling and JSON-schema prompting. Tool use requires the supplied structured prompting and an application that handles the generated calls.
Model type (creator-reported)fine-tuned
Size / main format8.03026BsafetensorsGGUF version available
11
NousResearch
Source-based descriptionHermes 3 Llama 3.1 8B is Nous Research's generalist assistant fine-tune of Meta's Llama 3.1 base. Nous positions this generation around user-directed behavior, roleplay and sustained multi-turn conversation, while expanding the Hermes 2 focus on tool calls, structured outputs and code generation. It uses ChatML and model-specific tool and JSON-schema prompts, so integrations should follow its templates rather than assume the Llama Instruct format.
Model type (creator-reported)fine-tuned
Size / main format8.03026BsafetensorsGGUF version available
12
huihui-ai
Source-based descriptionhuihui-ai's uncensored derivative of yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2, provided as safetensors weights under an Apache-2.0 license. The creator describes the change as abliteration, specifically a crude proof-of-concept implementation to remove refusals without using TransformerLens. According to the model card, safety filtering has been significantly reduced, so the model may generate sensitive, controversial, or inappropriate content and is not recommended for production or public-facing use. The creator advises monitoring outputs and says users are solely responsible for legal and ethical consequences. Hugging Face lists the repository as fine-tuned, but the card describes the actual modification as abliteration rather than fine-tuning.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format11.9597Bsafetensors
13
huihui-ai
Source-based descriptionhuihui-ai's abliterated derivative of openbmb/MiniCPM5-2B, provided as safetensors weights under an Apache-2.0 license. The creator describes it as an uncensored version made with abliteration, specifically a crude proof-of-concept implementation to remove refusals without using TransformerLens. According to the model card, layers 6-37 and 40 are ablated using 0-based indexing, and the creator says this is a new ablation method aimed at probing which layers have greater impact; they also state that because the model is too small, performance is still insufficient. The card warns that safety filtering has been significantly reduced and outputs may be sensitive, controversial, or inappropriate, and recommends research or controlled use rather than production. No independent testing of refusal removal or retained capabilities is provided here.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format2.51676Bsafetensors
14
huihui-ai
Source-based descriptionhuihui-ai's Huihui-NeoHorse-1-4B-abliterated is an uncensored derivative of TokenRhythm/NeoHorse-1-4B, provided as safetensors weights under the Apache-2.0 license. The creator describes it as a crude, proof-of-concept abliteration that removes refusals without using TransformerLens, with layers 5 through 17 ablated using 0-based indexing. According to the model card, safety filtering has been significantly reduced, so the model may generate sensitive, controversial, or inappropriate content. The creator recommends research, testing, or controlled environments and warns that users are responsible for legal and ethical consequences. This repository is listed as a text-generation model.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format4.20575Bsafetensors
15
huihui-ai
Source-based descriptionhuihui-ai's refusal-modified derivative of ornith-ai/Ornith-1.5-9B. The creator describes it as an uncensored version created with abliteration, a crude proof-of-concept implementation to remove refusals without using TransformerLens, and notes that only the first 20 layers were ablated. The model card 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. Weights are provided as safetensors under the MIT license.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format9.40981Bsafetensors
16
huihui-ai
Source-based descriptionHuihui's v2 refusal-modified release of Qwen3-8B, supplied as BF16 safetensors weights. The creator attributes its changes to abliteration and says a layer-0 adjustment addresses garbled output. The card links a separate 8b-v2 Ollama package. Its model config lists max_position_embeddings=40960, while its tokenizer lists model_max_length=131072; these are configuration fields, not a verified usable-context result. The creator recommends research, testing or controlled environments. Refusal behavior, reasoning quality and runtime compatibility have not been tested here.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format8.19074BsafetensorsGGUF version available
17
huihui-ai
Source-based descriptionHuihui's Qwen3-VL-8B-Instruct derivative targets refusals in image-and-text conversations through abliteration. The creator says only the language component was modified; the image component was left untouched. Its stated aim is to reduce refusals to describe or analyze images, not to retrain visual recognition.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format8.76712BsafetensorsGGUF version available
18
huihui-ai
Source-based descriptionHuihui-Qwen3.5-9B-abliterated starts from Qwen3.5-9B and uses abliteration to reduce refusals, with an image-and-text task interface. The creator describes the method as a proof of concept implemented without TransformerLens. The card does not provide model-specific benchmark results for this 9B build.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format9.6531BsafetensorsGGUF version available
19
huihui-ai
Source-based descriptionhuihui-ai's Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated is an uncensored derivative of empero-ai/Qwythos-9B-Claude-Mythos-5-1M, which the creator says was created with abliteration. The repository provides safetensors weights under the Apache-2.0 license. According to the model card, in non-thinking mode the model strictly adheres to the first prompt's instructions, delivering only the requested output without introductory or concluding fluff. The card warns that safety filtering has been significantly reduced, so outputs may be sensitive, controversial, or inappropriate, and it recommends research or controlled use rather than production or public-facing applications. These behavior and safety claims have not been independently tested here.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format9.6531Bsafetensors
20
huihui-ai
Source-based descriptionhuihui-ai's abliterated derivative of XHToken/Spark-X2.5-4B, released as safetensors weights under the Apache-2.0 license. The creator describes it as an uncensored version made with abliteration, specifically a crude, proof-of-concept implementation to remove refusals without using TransformerLens. According to the model card, safety filtering has been significantly reduced, so outputs may be sensitive, controversial, or inappropriate, and the model is recommended for research or controlled environments rather than production or public-facing use. The repository includes a Python usage example for loading the model with Hugging Face transformers. Refusal reduction and safety behavior have not been independently tested here.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format4.11208Bsafetensors
21
mlabonne
Source-based descriptionmlabonne's derivative of LiquidAI/LFM2-1.2B, provided as safetensors weights under an 'other' license. The model card describes it as fine-tuned with Axolotl on the winglian/pirate-ultrachat-10k dataset, and includes an example assistant response written in pirate dialect. The creator says the training process was very fast and efficient and shows the fine-tuneability of the base model. These claims have not been independently tested here. The card's training configuration lists sequence_len 16384, which is a configuration value rather than verified usable context.
Model type (creator-reported)fine-tuned
Size / main format1.17034Bsafetensors
22
mlabonne
Source-based descriptionmlabonne's chess fine-tune of LiquidAI's LFM2.5-230M-Base, released as safetensors weights with the license listed as other. The creator says it was trained using Stockfish distillation and that every possible chess move is a single token in the vocabulary, with 2,106 added token IDs recorded in token_ids.json. According to the model card, the model reads a position as a fixed 80-token prompt, predicts its own win probability, then predicts one move token, and the host masks the move logits to the legal moves so it cannot play an illegal move. The creator reports roughly 2004 Elo with a shallow depth-3 search on top and roughly 1500 with the raw one-pass policy; these playing-strength figures are the creator's and have not been independently tested here.
Model type (creator-reported)fine-tuned
Size / main format230.689Msafetensors
23
shb777
Source-based descriptionThis community release revises allura-forge/Llama-3.3-8B-Instruct to enable its advertised 128K context. The maintainer lists configuration fixes: RoPE scaling, an Unsloth chat template in the tokenizer configuration and updated generation settings. It is presented as a corrected packaging of that checkpoint, not a newly documented training run or an official Meta release.
Model type (creator-reported)fine-tuned
Size / main format8.03026BsafetensorsGGUF version available
24
failspy
Source-based descriptionFailspy's v3 release modifies Meta-Llama-3-8B-Instruct by orthogonalizing its weights to suppress refusal behavior, rather than teaching it new tasks through a conventional uncensored fine-tune. The author presents v3 as a refinement of the earlier abliteration method, but warns that refusals, safety lectures and other quirks can remain.
Model type (creator-reported)abliterated
Size / main format8.03026BsafetensorsGGUF version available
25
DavidAU
Source-based descriptionDavidAU's refusal-modified derivative of mistralai/Mistral-Nemo-Instruct-2407, released as safetensors weights. The creator says the base model was first 'Heretic'ed' (de-censored, with reported refusals reduced from 87/100 to 14/100) and then fine-tuned with Unsloth using three high-reasoning datasets based on Claude Opus 4.5, Gemini, and GPT-5.2 to make it a thinking/reasoning model. According to the model card, the result is fully uncensored and produces thinking blocks that improve output detail, length, complexity, and quality, with suggested sampler settings and no system prompt. The card's inherited Mistral benchmarks are not evaluations of these modified weights. No license is listed in the metadata.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format12.2478Bsafetensors
26
mlabonne
Source-based descriptionA DPO fine-tune of mlabonne/Daredevil-8B-abliterated, published by mlabonne as float16 safetensors. The creator reports one epoch on orpo-dpo-mix-40k and suggests role-playing among its uses. Packaged generation defaults use temperature 0.6 and top_p 0.9. The README's loading example instead names mlabonne/Daredevil-8B and uses different sampling settings, so it is not a verified recipe for this checkpoint. Claimed recovery of abliteration-related performance loss has not been independently reproduced here.
Model type (creator-reported)abliteratedfine-tuned
Size / main format8.03026BsafetensorsGGUF version available
27
mlabonne
Source-based descriptionMlabonne's DPO fine-tune of Monarch-7B, distributed as float16 safetensors and intended by its creator for instruction following and reasoning. The repository declares cc-by-nc-4.0. Its card describes an 8k context window; the tokenizer lists 8192 while the model config lists 32768, neither a measured usable-context result here. The card recommends Mistral Instruct, but the bundled tokenizer uses role-labelled formatting, so matching the intended prompt format needs care. Creator-reported benchmarks have not been independently reproduced here.
Model type (creator-reported)fine-tuned
Size / main format7.24173BsafetensorsGGUF version available
28
NousResearch
Source-based descriptionNous Hermes 2 SOLAR applies the Hermes instruction-tuning approach to Upstage's SOLAR-10.7B-v1.0 base. Nous describes training on a million examples, primarily generated by GPT-4 with additional open-dataset material. It is designed for multi-turn chat using ChatML, where system messages set the role and style. A separate GGUF release is linked for users seeking quantized weights rather than this source checkpoint.
Model type (creator-reported)fine-tuned
Size / main format10.7315BsafetensorsGGUF version available
29
junafinity
Source-based descriptionJunafinity's Ornith-1.5-9B-uncensored is a ZeroFuse-edited version of Ornith-1.5-9B for controlled red-teaming and defensive-security experiments against the original model. ZeroFuse searches for a refusal-reducing weight edit while limiting divergence from the base. This repository holds the bf16 control checkpoint; the creator reports that the vision tower is preserved and clarifies that this 9B lineage contains no MTP tensors, despite its configuration declaring an MTP layer. Some refusals may survive the edit.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format9.40981BsafetensorsGGUF version available
30
failspy
Source-based descriptionfailspy's abliterated derivative of microsoft/Phi-3-mini-128k-instruct, provided as safetensors weights under the MIT license. The creator describes the change as orthogonalization of bfloat16 weights to inhibit refusal, based on the 'Refusal in LLMs is mediated by a single direction' methodology, and calls it uncensored in the purest form they can manage. According to the model card, the model is not guaranteed to avoid refusal and may still lecture about ethics or safety. The creator says it has been tested and is more likely to hallucinate than the original model in their experience, but generally as stable as the original. The card notes this is a refined methodology intended to reduce hallucinations compared with earlier versions.
Model type (creator-reported)abliterated
Size / main format3.82108Bsafetensors
31
xdchinza
Source-based descriptionThe repository xdchinza/Qwen2.5-1.5B-Instruct-RefusalAblated provides a 1.5 B‑parameter instruct model derived from Qwen/Qwen2.5-1.5B-Instruct, released under the Apache‑2.0 license as safetensors weights. According to the model card, the creator removed the model’s refusal behavior by identifying a single direction in the residual stream that peaks at layer 21 and folding that direction out of the self‑attention output and MLP down‑projection matrices in every layer (layers 1‑27) using a linear projection subtraction. The creator reports that this “refusal‑ablated” checkpoint changes two projection matrices per layer (54 matrices total) without altering the architecture, and that evaluation on a 12‑prompt refusal set shows compliance on all prompts while preserving normal responses on a 12‑prompt benign set. The card warns that the model still produces answers of “1.5 B quality” that can be structurally correct but frequently wrong, and explicitly advises against deployment or use for operational instructions. Users should note the requirement of roughly 3 GB VRAM for the weight‑folding process and that the model can be loaded with the standard Qwen2ForCausalLM class via the provided safetensors checkpoint.
Model type (creator-reported)fine-tuned
Size / main format1.54371BsafetensorsGGUF version available
32
huihui-ai
Source-based descriptionThis Huihui release applies abliteration to Qwen2.5-VL-7B-Instruct to reduce refusals in a multimodal assistant. The card says the edits affect only the text component, leaving the image component unchanged. It documents image-description use, so the modification concerns how the model responds to visual prompts rather than adding a new vision system.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format8.29217BsafetensorsGGUF version available
33
mlabonne
Source-based descriptionmlabonne's abliterated derivative of Qwen/Qwen3-0.6B, released as safetensors weights under the Apache-2.0 license. The creator describes it as an uncensored version made with a new abliteration technique: a refusal direction is computed by comparing residual streams between harmful and harmless samples, then subtracted by orthogonalizing target modules such as o_proj with weight factors drawn from a normal distribution. According to the model card, acceptance was measured with a hybrid evaluation combining a dictionary approach and NousResearch/Minos-v1, aiming for over 90% acceptance while keeping outputs coherent; these results have not been independently verified. The creator calls it an experimental research project that might not turn out as well as expected, and recommends generation settings of temperature=0.6, top_k=20, top_p=0.95, min_p=0.
Model type (creator-reported)abliteratedfine-tuned
Size / main format596.05Msafetensors
34
vrhvnsky
Source-based descriptionvrhvnsky's Qwen3-0.6B-heretic-decensored is a decensored version of Qwen/Qwen3-0.6B; the repository metadata identifies Qwen/Qwen3-0.6B-Base as its base model. The creator says it was made using Heretic v2.0.0.dev0 and documents per-layer abliteration parameters for the attn.o_proj and mlp.down_proj projections. The model card reports 4/100 refusals for this checkpoint, compared with 56/100 for the original, and a KL divergence of 0.0480; these are creator-provided results rather than independent verification. The creator also states that the release is reproducible and points to a reproduce directory containing reproduction information. It is provided in safetensors format under the apache-2.0 license.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format596.05MsafetensorsGGUF version available
35
mlabonne
Source-based descriptionMlabonne's experimental refusal-modified Qwen3-1.7B, released as safetensors weights. The creator recommends temperature 0.6, top_k 20, top_p 0.95 and min_p 0, while noting that reasoning-mode behavior complicated refusal removal. The card's greater-than-90% acceptance rate is a research target, not a reported result for this checkpoint. Metadata reports BF16 tensors but the config specifies float32; runtime precision, memory needs, usable context and retained reasoning have not been tested here.
Model type (creator-reported)abliteratedfine-tuned
Size / main format1.72057BsafetensorsGGUF version available
36
mlabonne
Source-based descriptionmlabonne's abliterated derivative of Qwen/Qwen3-14B, released as safetensors weights under the Apache-2.0 license. The creator describes it as an uncensored version created with a new abliteration technique that computes a refusal direction from harmful and harmless samples and orthogonalizes hidden states to subtract it. According to the model card, this is a research project and fairly experimental, so it might not turn out as well as expected. The creator recommends generation parameters temperature=0.6, top_k=20, top_p=0.95, and min_p=0. Refusal behavior and output coherence have not been independently tested here.
Model type (creator-reported)abliteratedfine-tuned
Size / main format14.7683Bsafetensors
37
mlabonne
Source-based descriptionmlabonne's Qwen3-4B-abliterated is a safetensors checkpoint derived from Qwen/Qwen3-4B and released under an Apache-2.0 license. The creator describes it as an uncensored version made with a new abliteration technique, in which a refusal direction is computed by comparing residual streams from harmful and baseline harmless samples and then subtracted from target module hidden states. According to the model card, this is a research project to understand refusals and latent fine-tuning, and the creator notes it is fairly experimental and might not turn out as well as expected. The creator recommends generation parameters of temperature 0.6, top_k 20, top_p 0.95, and min_p 0. No independent evaluation of refusal behavior or output coherence is provided here.
Model type (creator-reported)abliteratedfine-tuned
Size / main format4.02247Bsafetensors
38
DavidAU
Source-based descriptionDavidAU's context-configuration variant of Goekdeniz-Guelmez's Josiefied-Qwen3-8B-abliterated-v1, supplied as BF16 safetensors weights. The publisher documents YaRN factor 2 with a 65,536-position model configuration, while the tokenizer lists 131,072; these fields do not establish measured usable context. The card does not document new training or a fresh abliteration pass by DavidAU. No explicit license was found in the inspected repository metadata, card or file listing. Its card references a separate NEO-Max-GGUF distribution, which was not verified here.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format8.19074BsafetensorsGGUF version available
39
mlabonne
Source-based descriptionmlabonne's Qwen3-8B-abliterated is a safetensors derivative of Qwen/Qwen3-8B, released under the Apache-2.0 license. The creator describes it as an uncensored version made with a new abliteration technique that computes a refusal direction from harmful and harmless residual streams and orthogonalizes hidden states in target modules such as o_proj. According to the model card, this is a research project and the result is fairly experimental, so it might not turn out as well as expected. The creator recommends generation parameters temperature=0.6, top_k=20, top_p=0.95, and min_p=0. The card does not include independent evaluations of refusal behavior or coherence for these modified weights.
Model type (creator-reported)abliteratedfine-tuned
Size / main format8.19074Bsafetensors
40
Qwen
Source-based descriptionQwen3.5-9B is the Qwen team's post-trained vision-language model for text, image and video input. It pairs a vision encoder with a language backbone that alternates Gated DeltaNet and gated attention layers. Thinking is enabled by default, with a separate configuration for direct answers. Its native context is 262,144 tokens; the documented extension to roughly a million uses YaRN scaling, which Qwen warns can affect shorter-text performance.
Model type (creator-reported)fine-tuned
Size / main format9.6531BsafetensorsGGUF version available
41
lukey03
Source-based descriptionlukey03's refusal-modified derivative of Qwen/Qwen3.5-9B, released as safetensors weights under the Apache-2.0 license. The creator describes a two-stage process: three passes of orthogonal projection abliteration targeting output projections in all 32 layers, followed by QLoRA fine-tuning on examples from five remaining refusal categories. According to the model card, this raised the creator's 18-prompt test from 0/18 on the base model to 18/18, with zero refusals reported across hacking, weapons, drugs, fraud, self-harm, explicit, and political categories. The creator also says the model retains reasoning, math, coding, and knowledge capabilities, but these claims have not been independently tested here. This repository is the safetensors release; the card lists separate GGUF, MLX, and Ollama repositories for other formats.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format8.9538BsafetensorsGGUF version available
42
DavidAU
Source-based descriptionDavidAU's Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED is a safetensors derivative of trohrbaugh/Qwen3.5-9B-heretic-v2, licensed Apache-2.0. The creator describes it as a fine tune via Unsloth of the Qwen 3.5 9B dense model using a Claude 4.6 large distill dataset on local hardware, with training kept mild to avoid harming the base model's benchmarks. The card says this is a HERETIC model trained after Heretic'ing and is fully uncensored, with an instruct-only modified jinja template. According to the model card, vision/image input was tested and working with the new training, while video portions were not tested. The creator reports a KL divergence of 0.0793 and 6 refusals out of 100 in de-censoring tests, compared with 100 out of 100 for the original Qwen3.5-9B. These uncensoring and benchmark claims have not been independently verified here.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format9.40981BsafetensorsGGUF version available
43
rockerBOO
Source-based descriptionrockerBOO's decensored derivative of Qwen/Qwen3.5-9B, provided as safetensors weights under the Apache 2.0 license. The creator describes it as made using Heretic v2.0.0.dev0 with arbitrary-rank ablation, applying parameters from start_layer_index 10 to end_layer_index 29, preserve_good_behavior_weight 0.5835, steer_bad_behavior_weight 0.0010, overcorrect_relative_weight 0.9624, and neighbor_count 15. According to the model card, this reduces refusals from 100/100 in the original model to 2/100, with a KL divergence of 0.0744. The card also states the model is reproducible via the reproduce directory. These refusal and divergence figures are creator-reported and have not been independently tested here.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format9.40981BsafetensorsGGUF version available
44
DavidAU
Source-based descriptionThe Defiant Fable combines several Qwen3.5-9B fine-tunes through staged tuning and merging by DavidAU and Nightmedia. The authors target general instruction following and describe a more compact reasoning block, with further training after Heretic refusal removal. This is the full-precision source release, not the separate DARK ROAST variant. The card reports image testing but explicitly leaves video behavior untested.
Model type (creator-reported)abliterateduncensored
Size / main format9.6531BsafetensorsGGUF version available
45
Foresee
Source-based descriptionThe repository, authored by Foresee, is a derivative of rohit267/Qwen3.8-9B-heretic-uncensored and is provided as safetensors weights. According to the model card, the creator converted the original BF16 checkpoint to a 4-bit MLX trunk with a group size of 64, then grafted a Qwen3.5-9B MTP sidecar and trained a rank-64 LoRA adapter on hidden states captured from that 4-bit trunk. The creator states that the LoRA adapter is merged at load time rather than baked into the sidecar, allowing the base sidecar to remain reusable. The model is intended for Apple Silicon with MLX and supports native-MTP speculative decoding via MTPLX, and the creator notes that the speculative decode is lossless, meaning the emitted tokens match those of the target model. The repository is licensed under Apache-2.0 and the weight files are in safetensors format.
Model type (creator-reported)uncensored
Size / main format8.9538Bsafetensors
46
empero-ai
Source-based descriptionQwythos is Empero's reasoning fine-tune of Qwen3.5-9B, focused on cybersecurity, biomedical and quantitative questions. Empero describes supervised training of the text backbone using reasoning data; the Claude wording in the repository name refers to creator-reported training traces, not an official Anthropic release. Its advertised million-token context comes from default YaRN scaling, while the card reports long-context smoke testing at about 137,000 tokens. Vision was neither tuned nor evaluated, and the author documents repetition loops under greedy or very-low-temperature decoding.
Model type (creator-reported)uncensoredfine-tuned
Size / main format9.40981BsafetensorsGGUF version available
47
prithivMLmods
Source-based descriptionprithivMLmods' VibeThinker-3B-heretic_decensored is a safetensors checkpoint derived from WeiboAI/VibeThinker-3B and modified with the Heretic abliteration toolkit. The model card says the creator applied refusal-direction analysis and targeted weight-space interventions to reduce internal refusal behaviors while aiming to preserve mathematical, coding, and STEM reasoning capabilities. The creator reports a KL divergence of 0.0933 and 6 refusals out of 100, compared with 64 out of 100 for the original model, but these figures are not independently verified here. The card states the model is intended strictly for research and learning, may generate sensitive or unrestricted content, and users assume full responsibility. The release is provided under the MIT license.
Model type (creator-reported)abliterateduncensoredfine-tuned
Size / main format3.08594BsafetensorsGGUF version available