Abliteration

Abliteration serves versions of open-weight models with reduced refusal behaviour over an OpenAI-compatible API. The LLMTR catalog carries two Abliteration models; both are called on /v1/chat/completions with the abliteration/ prefix, exactly as the catalog lists them. Both models read text only.

Model	Context	Input / Cache read / Output ($/1M)
abliteration/abliterated-model-large-v2	1,000,000	3.00 / 0.30 / 5.00
abliteration/abliterated-model-large	1,000,000	3.00 / 0.30 / 5.00

Repeated prompt prefixes can be served from a cache; those tokens bill at the cache read price in the table. A single answer produces at most 999,990 tokens; a larger max_tokens is refused with 400 invalid_request before the request is sent.

The provider has marked abliterated-model-large as a legacy release. Use abliteration/abliterated-model-large-v2 for new work.

Although refusals are reduced, requests involving child sexual exploitation or help with self-harm are blocked.

Chat

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "abliteration/abliterated-model-large-v2",
    "messages": [
      { "role": "system", "content": "Answer briefly and clearly." },
      { "role": "user", "content": "Summarise the obligations in this contract clause in three points." }
    ],
    "max_tokens": 4096
  }'

Both models support streaming, and JSON mode and JSON Schema through response_format.

temperature must be between 0 and 1, and top_p between 0.01 and 1. A value outside that range is refused with 400 invalid_request before the request is sent.

These models have no web search: a request that sends web_search_options is refused with 400 invalid_request.

Reasoning

Both models reason, and those tokens bill as output. Choose the level with the reasoning_effort field or with a suffix on the model id (abliteration/abliterated-model-large-v2:max).

Model	Levels	When no level is sent
abliteration/abliterated-model-large-v2	low, high, max	low
abliteration/abliterated-model-large	none, high, max	high

Reasoning cannot be turned off on abliteration/abliterated-model-large-v2. A request that sends "reasoning_effort": "none", "reasoning": false or :fast is refused with 400 invalid_request; use low for the least reasoning.

To turn reasoning off on abliteration/abliterated-model-large, send "reasoning_effort": "none" or "reasoning": false, or append :none to the model id. With it off, no reasoning tokens are produced or billed.

A level that is not listed is refused with 400 invalid_request. If an answer returns with finish_reason: "length", raise max_tokens and retry.

Tool calling

Both models support tool/function calling; the model chooses the tool itself. The tool_choice field is honoured only as auto. A request that sends required or a named function is refused with 400 invalid_request; say in the prompt when the tool should be used.

A request that sends tool_choice: "none" with tools attached is refused with 400 unsupported_input. To turn tool use off, do not send the tools field at all.

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "abliteration/abliterated-model-large-v2",
    "messages": [{ "role": "user", "content": "What is the weather in Istanbul?" }],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_weather",
          "description": "Get the current weather for a city",
          "parameters": {
            "type": "object",
            "properties": { "city": { "type": "string" } },
            "required": ["city"]
          }
        }
      }
    ],
    "max_tokens": 1024
  }'

On the second turn send back the model's message with the tool call and the tool's result (role: "tool"); you do not need to send the reasoning text back.