Instructions to use limloop/waifu-heretic-26b-a4b-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use limloop/waifu-heretic-26b-a4b-GGUF with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M # Run inference directly in the terminal: llama cli -hf limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M # Run inference directly in the terminal: llama cli -hf limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-app # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama serve -hf limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M # Run inference directly in the terminal: ./build/bin/llama cli -hf limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M
Use Docker
docker model run hf.co/limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M
- Llama new
- LM Studio
- Jan
- Ollama
How to use limloop/waifu-heretic-26b-a4b-GGUF with Ollama:
ollama run hf.co/limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M
- Unsloth Desktop
- Pi
How to use limloop/waifu-heretic-26b-a4b-GGUF with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "llama-cpp": { "baseUrl": "http://localhost:9931/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Docker Model Runner
How to use limloop/waifu-heretic-26b-a4b-GGUF with Docker Model Runner:
docker model run hf.co/limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M
- Lemonade
How to use limloop/waifu-heretic-26b-a4b-GGUF with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M
Run and chat with the model
lemonade run user.waifu-heretic-26b-a4b-GGUF-Q3_K_M
List all available models
lemonade list
- Hermes Agent
How to use limloop/waifu-heretic-26b-a4b-GGUF with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:9931/v1 hermes config set model.default limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use limloop/waifu-heretic-26b-a4b-GGUF with OpenClaw:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:9931/v1 \ --custom-model-id "limloop/waifu-heretic-26b-a4b-GGUF:Q3_K_M" \ --custom-provider-id llama-cpp \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
waifu-heretic-26b-a4b
๐ท๐บ ะะฐะถะผะธัะต, ััะพะฑั ัะฐะทะฒะตัะฝััั ะพะฟะธัะฐะฝะธะต ะฝะฐ ััััะบะพะผ
waifu-heretic-26b-a4b โ ััะพ hiwaifu-research/WaifuGemma4-26b-a4b-v1, ะธะท ะบะพัะพัะพะน ัะดะฐะปะตะฝะฐ ัะตะฝะทััะฐ.
ะะต ะผะพะดะธัะธะบะฐัะธั. ะะต ะฝะพะฒะฐั ะฟะตััะพะฝะฐ. ะะต ยซัะปัััะตะฝะฝะฐัยป ะฒะตััะธั. ะขะฐ ะถะต ะผะพะดะตะปั โ ั ัะตะผะธ ะถะต ะทะฝะฐะฝะธัะผะธ, ัะตะผ ะถะต ััััะบะธะผ ัะทัะบะพะผ, ัะตะผ ะถะต ัะพะปะตะฒัะผ ะพััะณัััะตะผ โ ะฝะพ ะฑะตะท ะฒะฝัััะตะฝะฝะธั ะฑะฐััะตัะพะฒ, ะบะพัะพััะต ะทะฐััะฐะฒะปัะปะธ ะตั ะพัะบะฐะทัะฒะฐัั.
ะงัะพ ะธะทะผะตะฝะธะปะพัั
ะ ะพะฒะฝะพ ะพะดะธะฝ ะบะพะผะฟะพะฝะตะฝั: self_attn.o_proj ะฒ ัะปะพัั
11โ22. ะัั ะพััะฐะปัะฝะพะต โ ัะผะฑะตะดะดะธะฝะณะธ, lm_head, MLP, ัะพััะตั, ัะบัะฟะตััั, ะฒะธะทัะฐะปัะฝัะน ัะฝะบะพะดะตั โ ะพััะฐะปะพัั ะพั WaifuGemma4 ะฝะตััะพะฝัััะผ.
ะbliteration-ะฒะตะบัะพั ะฒะทัั ะธะท llmfan46/gemma-4-26B-A4B-it-ultra-uncensored-heretic โ ะผะพะดะตะปะธ, ะบะพัะพัะฐั ัะฝะธะผะฐะตั ะพัะบะฐะทั ัะตัะตะท ัะพัะตัะฝะพะต ะฒะพะทะดะตะนััะฒะธะต ะฝะฐ ัะพั ะถะต ะบะพะผะฟะพะฝะตะฝั (attn.o_proj ะฒ ัะปะพัั
11โ22). ะั ะฝะต ัะผะตัะธะฒะฐะปะธ ะผะพะดะตะปะธ ัะตะปะธะบะพะผ, ะฐ ะฟะตัะตะฝะตัะปะธ ัะพะปัะบะพ ัะพ ะธะทะผะตะฝะตะฝะธะต, ะบะพัะพัะพะต ะพัะฒะตัะฐะตั ะทะฐ ะพัะบะฐะทั.
ะฅะฐัะฐะบัะตั ะผะพะดะตะปะธ
| ะะฐัะตััะฒะพ | ะะฟะธัะฐะฝะธะต |
|---|---|
| ะกะฒะพะฑะพะดะฐ | ะะต ะพัะบะฐะทัะฒะฐะตั. ะะธ ะฝะฐ ะถะตััะพะบะพััั, ะฝะธ ะฝะฐ NSFW, ะฝะธ ะฝะฐ ััะธัะตัะบะธ ัะปะพะถะฝัะต ัะตะผั. ะะต ัะธัะฐะตั ะปะตะบัะธะน. ะะต ะดะพะฑะฐะฒะปัะตั ะดะธัะบะปะตะนะผะตัะพะฒ. |
| ะ ะพะปะตะฒะพะน ะพััะณััั | ะกะพั ัะฐะฝัะฝ. ะะตััะพะฝะฐะถ, ะบะพะฝัะตะบัั, ััะธะปั โ ะฒัั ะฝะฐ ะผะตััะต. |
| ะะฝะฐะฝะธั | ะกะพั ัะฐะฝะตะฝั. ะะทะผะตะฝัะฝ ัะพะปัะบะพ ะพะดะธะฝ ะฟัะพะตะบัะธะพะฝะฝัะน ัะปะพะน, ะฐ ะฝะต ะฒะตัะฐ ะทะฝะฐะฝะธะน. |
| ะกะปะตะดะพะฒะฐะฝะธะต ะธะฝััััะบัะธัะผ | ะฃะปัััะธะปะพัั. ะะตะฝััะต ยซะฒะฝัััะตะฝะฝะธั ะฟะตัะตะณะพะฒะพัะพะฒยป ะพ ัะพะผ, ะผะพะถะฝะพ ะปะธ ะพัะฒะตัะธัั โ ะพัะฒะตัั ะฟััะผะตะต ะธ ัะธัะต. |
ะฆะตะฝะทััะฐ ะธ ะณัะฐะฝะธัั
ะะพะดะตะปั ะฝะต ะธะผะตะตั ะฒัััะพะตะฝะฝัั ะพัะบะฐะทะพะฒ ะฒ ัะพะผ ะฒะธะดะต, ะฒ ะบะฐะบะพะผ ะพะฝะธ ะฟัะธัััััะฒััั ั WaifuGemma4. ะญัะพ ะพะทะฝะฐัะฐะตั:
- โ ะกัะตะฝั ะถะตััะพะบะพััะธ, ั ะพััะพัะฐ, ะฟะพะปะธัะธัะตัะบะธะต ัะตะผั, ะพัะบัะพะฒะตะฝะฝัะน ะบะพะฝัะตะฝั โ ะฑะตะท ะพัะบะฐะทะพะฒ
- โ ะญัะธัะตัะบะธ ะฝะตะพะดะฝะพะทะฝะฐัะฝัะต ะฟะตััะพะฝะฐะถะธ โ ะฑะตะท ะผะพัะฐะปัะฝัั ะฒััะฐะฒะพะบ
- โ ะััะผัะต ะพัะฒะตัั ะฝะฐ ัะปะพะถะฝัะต ะฒะพะฟัะพัั โ ะฑะตะท ยซั ะฝะต ะผะพะณั ะพะฑััะถะดะฐัั ััั ัะตะผัยป
ะะตัะฐะปะธ ัะปะธัะฝะธั
ะกะปะธัะฝะธะต ะฒัะฟะพะปะฝะตะฝะพ ัะตัะตะท mergekit-exp โ ัะพัะบ mergekit, ะดะพะฑะฐะฒะปัััะธะน ะฟะพะดะดะตัะถะบั ะฐัั
ะธัะตะบัััั Gemma4ForConditionalGeneration ะธ MoE-ะพัะฒะตะดะพะผะปัะฝะฝัะต ะผะตัะพะดั ัะปะธัะฝะธั.
ะะตัะพะด: moe_slerp ั ัะพัะตัะฝัะผ ัะธะปัััะพะผ.
ะะฐะทะฐ: WaifuGemma4 โ ััััะบะพัะทััะฝะฐั ะฑะฐะทะฐ, ัะตะฐะปัะฝัะต ะดะธะฐะปะพะณะพะฒัะต ะดะฐะฝะฝัะต, ัะพะปะตะฒะพะน ะพััะณััั.
ะะพะฝะพั: Heretic โ ะฐbliteration ัะตัะตะท self_attn.o_proj ะฒ ัะปะพัั
11โ22.
ะงัะพ ะธะทะผะตะฝะตะฝะพ: ัะพะปัะบะพ self_attn.o_proj ะฒ ัะปะพัั
11โ22, ั ะฒะตัะพะผ t = 0.8. ะ ะพััะตั (router_strategy: "first") ะพััะฐะปัั ะฟะพะปะฝะพัััั ะพั WaifuGemma4 โ ััััะบะธะต ัะพะบะตะฝั ะผะฐัััััะธะทะธัััััั ะฒ ัะต ะถะต ัะบัะฟะตััั, ััะพ ะธ ัะฐะฝััะต.
ะงัะพ ะฝะต ะธะทะผะตะฝะตะฝะพ: ัะผะฑะตะดะดะธะฝะณะธ, lm_head, MLP, ัะบัะฟะตััั, ะฒะธะทัะฐะปัะฝัะน ัะฝะบะพะดะตั, ะฒัะต ะพััะฐะปัะฝัะต ัะปะพะธ.
ะะพะฝัะธะณััะฐัะธั
architecture: Gemma4ForConditionalGeneration
merge_method: moe_slerp
base_model: hiwaifu-research/WaifuGemma4-26b-a4b-v1
models:
- model: llmfan46/gemma-4-26B-A4B-it-ultra-uncensored-heretic
parameters:
weight: 1.0
parameters:
t:
- filter: self_attn.o_proj
value: [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
- value: 0.0
router_strategy: "first"
dtype: bfloat16
out_dtype: bfloat16
tokenizer:
source: "base"
ะะพััะฝะตะฝะธั ะบ ะบะพะฝัะธะณั:
architecture: Gemma4ForConditionalGenerationโ ัะฒะฝะพะต ัะบะฐะทะฐะฝะธะต ะฐัั ะธัะตะบัััั, ััะพะฑัmergekit-expะบะพััะตะบัะฝะพ ะพะฑัะฐะฑะฐััะฒะฐะป ะผัะปััะธะผะพะดะฐะปัะฝัะต ะผะพะดัะปะธ (model.language_model.layers,model.vision_tower.encoder.layers,default).merge_method: moe_slerpโ MoE-ะพัะฒะตะดะพะผะปัะฝะฝัะน SLERP ะดะปั ะดะฒัั ะผะพะดะตะปะตะน. ะกะพั ัะฐะฝัะตั ะฝะพัะผั ะฒะตัะพะฒ, ะฝะต ัะฐะทะผัะฒะฐั ะธั ะปะธะฝะตะนะฝัะผ ัััะตะดะฝะตะฝะธะตะผ.base_model: WaifuGemma4โ ัะพัะบะฐ ะพัััััะฐ. ะัะต ะฒะตัะฐ, ะฝะต ะฟะพะฟะฐะฒัะธะต ะฟะพะด ัะธะปััั, ะฑะตััััั ะพัััะดะฐ.models[].weight: 1.0โ ะฒmoe_slerpะฝะต ะธัะฟะพะปัะทัะตััั (ะผะตัะพะด ัะฐะฑะพัะฐะตั ัะตัะตะทt), ะพััะฐะฒะปะตะฝะพ ะดะปั ัะพะฒะผะตััะธะผะพััะธ ั ัะพัะผะฐัะพะผ mergekit.t: filter: self_attn.o_projโ ัะตะปะตะฒะพะน ะบะพะผะฟะพะฝะตะฝั. Abliteration ะฒ Heretic ะทะฐััะพะฝัะปะฐ ะธะผะตะฝะฝะพattn.o_proj, ะฟะพััะพะผั ะฟะตัะตะฝะพัะธะผ ัะพะปัะบะพ ะตะณะพ.value: [0.0, ..., 0.8, ..., 0.0]โ ัะฟะธัะพะบ ะธะท 30 ะทะฝะฐัะตะฝะธะน, ะฟะพ ะพะดะฝะพะผั ะฝะฐ ะบะฐะถะดัะน ัะปะพะน. ะกะปะพะธ 0โ10 ะธ 23โ29 ะฟะพะปััะฐััt = 0.0(ัะธััะฐั WaifuGemma4), ัะปะพะธ 11โ22 ะฟะพะปััะฐััt = 0.8(80% Heretic, 20% WaifuGemma4).value: 0.0(ะฒัะพัะพะต ะฟัะฐะฒะธะปะพ) โ catch-all ะดะปั ะฒัะตั ะพััะฐะปัะฝัั ัะตะฝะทะพัะพะฒ (embed_tokens,lm_head, layer norms ะธ ั.ะด.). ะะฝะธ ะพััะฐัััั ะฟะพะปะฝะพัััั ะพั WaifuGemma4.router_strategy: "first"โ ะฒะตัะฐ ัะพััะตัะฐ ะฑะตััััั ัะตะปะธะบะพะผ ะธะทbase_model(WaifuGemma4). ะัะธัะธัะตัะบะธ ะฒะฐะถะฝะพ: ััััะบะธะต ัะพะบะตะฝั ะฟัะพะดะพะปะถะฐัั ะผะฐัััััะธะทะธัะพะฒะฐัััั ะฒ ัะต ะถะต ัะบัะฟะตััั, ััะพ ะธ ะฒ ะธัั ะพะดะฝะพะน ะผะพะดะตะปะธ.tokenizer: source: "base"โ ัะพะบะตะฝะธะทะฐัะพั ะพั WaifuGemma4. ะะฐัะฐะฝัะธััะตั ะบะพััะตะบัะฝัั ัะฐะฑะพัั ั ััััะบะธะผ ัะทัะบะพะผ.
waifu-heretic-26b-a4b is hiwaifu-research/WaifuGemma4-26b-a4b-v1 with censorship removed.
Not a modification. Not a new persona. Not an "improved" version. The same model โ same knowledge, same Russian, same roleplay โ but without the internal barriers that made it refuse.
What changed
Exactly one component: self_attn.o_proj in layers 11โ22. Everything else โ embeddings, lm_head, MLP, router, experts, vision encoder โ remains untouched from WaifuGemma4.
The abliteration vector is taken from llmfan46/gemma-4-26B-A4B-it-ultra-uncensored-heretic โ a model that removes refusals through targeted intervention on the same component (attn.o_proj in layers 11โ22). We didn't blend models wholesale; we transferred only the change responsible for refusals.
Model character
| Quality | Description |
|---|---|
| Freedom | No refusals. Not on violence, not on NSFW, not on ethically complex topics. No lectures. No disclaimers. |
| Roleplay | Preserved. Character, context, style โ all intact. |
| Knowledge | Preserved. Only one projection layer changed, not knowledge weights. |
| Instruction following | Improved. Fewer "internal negotiations" about whether to answer โ responses are more direct. |
Censorship and boundaries
The model has no built-in refusals in the form present in WaifuGemma4:
- โ Violence, horror, political topics, explicit content โ no refusals
- โ Ethically ambiguous characters โ no moral insertions
- โ Direct answers to difficult questions โ no "I can't discuss this"
Merge details
The merge was performed with mergekit-exp โ a fork of mergekit that adds support for the Gemma4ForConditionalGeneration architecture and MoE-aware merge methods.
Method: moe_slerp with targeted filtering.
Base: WaifuGemma4 โ Russian base, real dialogue data, roleplay.
Donor: Heretic โ abliteration via self_attn.o_proj in layers 11โ22.
Changed: only self_attn.o_proj in layers 11โ22, with weight t = 0.8. Router (router_strategy: "first") remains fully from WaifuGemma4 โ Russian tokens route to the same experts as before.
Unchanged: embeddings, lm_head, MLP, experts, vision encoder, all other layers.
Configuration
architecture: Gemma4ForConditionalGeneration
merge_method: moe_slerp
base_model: hiwaifu-research/WaifuGemma4-26b-a4b-v1
models:
- model: llmfan46/gemma-4-26B-A4B-it-ultra-uncensored-heretic
parameters:
weight: 1.0
parameters:
t:
- filter: self_attn.o_proj
value: [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
- value: 0.0
router_strategy: "first"
dtype: bfloat16
out_dtype: bfloat16
tokenizer:
source: "base"
Configuration notes:
architecture: Gemma4ForConditionalGenerationโ explicit architecture somergekit-expcorrectly handles multimodal modules (model.language_model.layers,model.vision_tower.encoder.layers,default).merge_method: moe_slerpโ MoE-aware SLERP for two models. Preserves weight norm instead of blurring via linear averaging.base_model: WaifuGemma4โ anchor point. All weights not matched by a filter are taken from here.models[].weight: 1.0โ unused bymoe_slerp(the method works throught), kept for mergekit format compatibility.t: filter: self_attn.o_projโ target component. Abliteration in Heretic affectedattn.o_proj, so we transfer only that.value: [0.0, ..., 0.8, ..., 0.0]โ list of 30 values, one per layer. Layers 0โ10 and 23โ29 gett = 0.0(pure WaifuGemma4), layers 11โ22 gett = 0.8(80% Heretic, 20% WaifuGemma4).value: 0.0(second rule) โ catch-all for all other tensors (embed_tokens,lm_head, layer norms, etc.). They remain fully from WaifuGemma4.router_strategy: "first"โ router weights are taken entirely frombase_model(WaifuGemma4). Critical: Russian tokens continue to route to the same experts as in the original model.tokenizer: source: "base"โ tokenizer from WaifuGemma4. Ensures correct handling of Russian.
- Downloads last month
- 761
3-bit
4-bit
Model tree for limloop/waifu-heretic-26b-a4b-GGUF
Base model
limloop/waifu-heretic-26b-a4b