Abliterated Large V2 API

On Zurelay, abliterated-large-v2 costs $2.90 per 1M input tokens e $4.85 per 1M output tokens. Abliteration AI's list price is $3.00 input e $5.00 output per 1M, so you save 3%. Reasoning tokens bill as output, one rate applies at every prompt length, and failed requests are never billed.

Abliteration removes a model's refusals by editing its weights rather than retraining it. It compares the model's internal activations on prompts it refuses e prompts it answers, finds the direction linked to refusing e suppresses it, a method described in 2024 research by Arditi and colleagues. Abliteration AI's documentation links Large V2 to Z.ai's GLM-5.3, and notes that other abilities can shift along with the refusals, so evaluate it on your own tasks.

Work that mainstream models often refuse or water down: authorized security research e red-teaming, explanations of penetration-testing techniques, testing your own AI systems with adversarial prompts, fiction with dark themes, e frank answers on controversial or sensitive topics without moralizing.

Yes. Abliteration removes the model's refusals, not the rules. Zurelay's Terms de Service (https://zurelay.com/legal/terms) apply to everything sent to Abliterated Large V2, including the acceptable use rules in section 7, e you are responsible for how you use its output, as with any model on Zurelay.

Install the official openai package, set base_url to https://api.zurelay.com/v1 e use your Zurelay API key. Then send a Chat Completions request with model set to "abliterated-large-v2". Zurelay also accepts abliterated-model-large-v2 e abliteration-ai/abliterated-model-large-v2.

It holds 1,000,000 tokens of context, e in our tests it found one fact buried in a 112K-token prompt in 11 seconds. Reasoning e the answer share the output budget you set with max_tokens, so leave room for both.

No. It takes text only, as on Abliteration AI's own API. A request that includes imagemms or files is turned away with a clear error before anything is billed.

It always reasons a little before it answers. Set reasoning_effort to low, high or max; max is the default. Reasoning can't be switched off, e a request for none runs em low. Abliteration AI's documentation says the same of its own API.

Yes, with one catch. The model reasons before it answers, e a stop sequence that also turns up in its reasoning ends the response early, sometimes before any answer text. Use a stop sequence that only belongs in the answer, such as a distinctive end marker, e it works as expected. max_tokens and the usual sampling settings work too.

Yes. Pass tools as you would to any OpenAI model; it can call several at once, also while streaming, e reads tool results. tool_choice "required", a tool forced by name e "none" all work. For JSON, send response_format with json_object or json_schema and the answer comes back as JSON. logprobs isn't supported.

Yes. Requests for abliterated-large-v2 run on Abliteration AI's Abliterated Large V2, e Zurelay doesn't swap in a smaller or substitute model. Asked what it is, it names itself Abliterated Large V2, made by Abliteration AI, e it keeps any persona your system prompt gives it. Automatic smart routing has no alternative for it, so a request never quietly falls back to a model that still has its refusals; only a fallback you set yourself, in a key's custom rules or a request's models list, can answer instead, and the x-zurelay-model header names the model that answered. Responses are sampled, so wording varies between runs. Zurelay is independent and is not affiliated with or endorsed by Abliteration AI.