Abliteration

Use Abliteration through LiteLLM with the abliteration/ model prefix. LiteLLM reads ABLITERATION_API_KEY and sends requests to https://api.abliteration.ai/v1.

Get started

Create an API key in your Abliteration account and set ABLITERATION_API_KEY in your environment or secret manager. Install litellm for the Python SDK, or litellm[proxy] for the gateway.

export ABLITERATION_API_KEY="YOUR_ABLITERATION_API_KEY"
pip install litellm

from litellm import completion

response = completion(
    model="abliteration/abliterated-model-large-v2",
    messages=[{"role": "user", "content": "Reply with hello."}],
    reasoning_effort="low",
)
print(response.choices[0].message.content)

A successful request returns an assistant reply and token usage. You do not need to set api_base.

Models and pricing

Prices below are USD per million tokens, checked against the published pricing and live model catalog on October 26, 2026. Context is the combined input and output budget, so the full output limit is available only when the prompt fits in the remaining context. Prefix each model ID with abliteration/ in LiteLLM.

Model ID	Context	Max output	Input	Cached input	Output
abliterated-model	262,144	262,134	$1	$0.1	$3
abliterated-model-large	1,000,000	999,990	$3	$0.3	$5
abliterated-model-large-v2	1,000,000	999,990	$3	$0.3	$5
abliterated-model-large-flash	1,048,576	262,144	$1	$0.1	$2

abliterated-model-large is deprecated but remains available for existing integrations. Use abliterated-model-large-v2 for new text workloads. abliterated-model and abliterated-model-large-flash accept images. The two other models accept text only.

Gateway setup

Install the proxy with pip install 'litellm[proxy]'. Set ABLITERATION_API_KEY on the gateway process, and set LITELLM_MASTER_KEY to a strong key for gateway clients. The provider key authenticates upstream requests. The gateway key authenticates your client to LiteLLM.

Add this deployment to your existing configuration, or save it as config.yaml for a new gateway.

model_list:
  - model_name: abliteration-chat
    litellm_params:
      model: abliteration/abliterated-model-large-v2
      api_key: os.environ/ABLITERATION_API_KEY

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY

litellm --config config.yaml --port 4000

In a second terminal with LITELLM_MASTER_KEY set, send a request.

curl http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "abliteration-chat",
    "messages": [{"role": "user", "content": "Reply with hello."}],
    "reasoning_effort": "low"
  }'

Expect HTTP 200 with an assistant message in choices[0].message.content and counts in usage.

Streaming and prompt caching

Set stream=True and request usage in the final chunk.

from litellm import completion

stream = completion(
    model="abliteration/abliterated-model-large-v2",
    messages=[{"role": "user", "content": "Reply with hello."}],
    reasoning_effort="low",
    stream=True,
    stream_options={"include_usage": True},
)
for chunk in stream:
    if chunk.choices:
        print(chunk.choices[0].delta.content or "", end="")
    if getattr(chunk, "usage", None):
        print(chunk.usage)

LiteLLM uses usage.prompt_tokens_details.cached_tokens to charge reused input at the cached-input rate. Uncached input uses the standard input rate. A repeated prompt does not guarantee a cache hit.

Abliteration caches prompt prefixes automatically. Cache creation uses the standard input rate. Keep stable instructions and tool definitions at the start of the prompt. Pass a stable prompt_cache_key per conversation to improve reuse. Cache retention is not guaranteed, and prompt_cache_retention is unsupported. See Abliteration prompt caching.

Tools, structured output, and reasoning

Chat Completions supports tools, tool_choice, and response_format. Use LiteLLM function calling for tool-result continuations and structured outputs for JSON Schema examples.

Use reasoning_effort="low", "high", or "max" to select reasoning depth. none disables reasoning on all models except Large v2, where it selects low reasoning and hides the trace. LiteLLM excludes none from Large v2's advertised reasoning choices because this model cannot turn reasoning off. Reasoning text, when returned, is available as message.reasoning_content. On native Responses, abliterated-model accepts effort through xhigh but rejects max. See Abliteration reasoning.

Responses API

LiteLLM translates responses() requests into Chat Completions for this provider. The example below uses the same model and token rates as completion(). Abliteration also publishes a native Responses API, but this integration keeps LiteLLM's existing translation route.

from litellm import responses

response = responses(
    model="abliteration/abliterated-model-large-v2",
    input="Reply with hello.",
    reasoning={"effort": "low"},
    store=False,
)
print(response.output)

Gateway clients can send the same request to /v1/responses with the configured deployment name.

Anthropic Messages API

LiteLLM translates Anthropic Messages requests into Chat Completions for this provider. Use the same gateway deployment with an Anthropic client, or call LiteLLM directly. Anthropic Messages requires max_tokens. This example uses the provider's published output maximum for Large v2.

import asyncio
import litellm

async def main():
    response = await litellm.anthropic.messages.acreate(
        model="abliteration/abliterated-model-large-v2",
        messages=[{"role": "user", "content": "Reply with hello."}],
        max_tokens=999990,
        output_config={"effort": "low"},
    )
    print(response["content"])

asyncio.run(main())

Web search is a separate provider charge of $8 per 1,000 searches plus token usage. The model token rates above do not include that charge or account-specific subscription discounts. See Abliteration pricing.

Troubleshooting

For 401, check the key in the gateway process environment. For 402, check provider credits. For 404, compare the unprefixed model ID with GET https://api.abliteration.ai/v1/models. For 429, honor the provider's retry guidance. A working model catalog does not guarantee that every listed model is currently available for inference.