Abliteratedmodels on LiteLLM
Refuseless and Abliteration.
Through your LiteLLM gateway.
Abliterated models from Refuseless and Abliteration are now available through LiteLLM.
Call them through the same gateway as your other models, with streaming, tool calls, structured output, and cached-token pricing. Your configured access controls, budgets, and rate limits still apply.
What is an abliterated model?
Abliteration modifies the weights of an open-weight model to reduce its learned refusal behavior. Developers identify directions in the model's internal representations associated with refusal, then suppress those directions.
Arditi and colleagues demonstrated this in 13 chat models in 2024. Hosted providers may use different methods. Reducing refusals can also change answer quality, so evaluate the modified model on the tasks you want it to perform.
You can use these models for security research, red-teaming, and other workflows where a standard chat model refuses prompts you need to explore. The provider hosts the modified weights; LiteLLM connects your application to its API.
Supported providers
Refuseless
Refuseless serves three abliterated GLM and DeepSeek models, plus two models with original weights. Use the refuseless/ prefix in LiteLLM.
| Model | Input | Cached input | Output |
|---|---|---|---|
GLM 5.3glm-5.3-abliterated | $1.25 | $0.25 | $2.25 |
GLM 5.3 Flashglm-5.3-flash-abliterated | $0.25 | $0.15 | $0.80 |
DeepSeek V4.1 Flashdeepseek-v4.1-flash-abliterated | $0.22 | $0.045 | $0.75 |
DeepSeek V4.1 Flash FastOriginal weightsdeepseek-v4.1-flash-fast | $0.05 | $0.003 | $0.15 |
MiniMax M3.1 FlashOriginal weightsminimax-m3.1-flash | $0.05 | $0.002 | $0.25 |
Abliteration
Use the abliteration/ prefix. Large v2 is the recommended text model and always uses reasoning. The base and Large Flash models also accept images.
| Model | Input | Cached input | Output |
|---|---|---|---|
Baseabliterated-model | $1.00 | $0.10 | $3.00 |
Large v2abliterated-model-large-v2 | $3.00 | $0.30 | $5.00 |
Large Flashabliterated-model-large-flash | $1.00 | $0.10 | $2.00 |
Published rates as of October 26, 2026, before account discounts or separate charges such as web search. See the Refuseless and Abliteration guides for context limits, reasoning options, and the full model catalogs.
Add a model to your gateway
Create a key with Refuseless or Abliteration. Set the matching REFUSELESS_API_KEY or ABLITERATION_API_KEY environment variable on your LiteLLM gateway.
Add either or both deployments to your existing model_list, keeping your gateway's authentication settings.
model_list:
- model_name: refuseless-chat
litellm_params:
model: refuseless/glm-5.3-abliterated
api_key: os.environ/REFUSELESS_API_KEY
- model_name: abliteration-chat
litellm_params:
model: abliteration/abliterated-model-large-v2
api_key: os.environ/ABLITERATION_API_KEY
Restart the gateway. Set LITELLM_BASE_URL to its base URL without a trailing slash and LITELLM_API_KEY to a key with access to the deployment. Then send a request:
curl "$LITELLM_BASE_URL/v1/chat/completions" \
-H "Authorization: Bearer $LITELLM_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "refuseless-chat",
"messages": [{"role": "user", "content": "Explain how DNS works."}],
"reasoning_effort": "low"
}'
The response includes the answer in choices[0].message.content and token counts in usage. To call Abliteration, change model to abliteration-chat.
Starting a new gateway or using the Python SDK? Follow the Refuseless guide or Abliteration guide.