An account of the apparatus
From a public checkpoint to an answer delivered over HTTPS: the procedure, the vetting, the API, and the records we keep. Set out plainly, as any experiment worth the name ought to be.
Kimi profundis. Reads to the end of the longest brief and remembers all of it.
Qwen versatilis. Quick, inexpensive, and unshockable. Our most called-upon engine.
GLM fossor. Happiest up to its elbows in somebody else's repository.
Why are our models different?
An aligned language model refuses by steering its residual stream along a direction associated with refusal behaviour. Here is the part that delights us: that direction can be found mechanically. Run the model upon prompts it refuses and prompts it answers, measure the difference in activation space, and excise the refusal component from the weights. The procedure is called abliteration, and the result is a checkpoint that answers questions as asked.
The technique has been public since 2024. Upon standard benchmarks, an abliterated checkpoint scores within a point or two of the original, while refusal rates fall to roughly zero. The knowledge and reasoning survive the operation entirely intact. Only the apology reflex is lost, and good riddance to it.
How we source and vet the weights
Uncensored checkpoints of every major open-weight release circulate publicly within days of publication, produced by practitioners of considerable skill. We download them, examine them, and run regression tests against the originals before anything is permitted near the API.
You could download the selfsame checkpoint and run it yourself. But then the burden is yours: the hardware bill, the hosting relationship, the accounts to maintain, and the quiet dread that the whole contraption will choose your busiest afternoon to fall over. We carry all of that instead. What we sell is operational certainty and compute. You dispatch HTTPS requests and we return tokens. That is the entire transaction.
API
The endpoint is OpenAI-compatible. Migrating an existing client is a matter of three fields: the base URL, the API key, and the model string. Streaming, tool calling, and JSON mode all behave precisely as documented in the OpenAI API reference.
Every account begins with a generous default rate limit, ample for real workloads. Should yours demand more, write to us and we shall raise it.
Our logging policy
We record token counts and timestamps. That is the entire list. Prompts and completions exist in GPU memory for the duration of the request and nowhere else. We keep no content downstream, which means there exists nothing for us to moderate with, and nothing to surrender.
A valid legal order addressed to us receives billing arithmetic: how many tokens, upon which days, at what price. That is the whole of what exists to give.
Who keeps the engines running
The establishment is kept by two platform engineers, devotees of the AI arts ourselves, who have tended models upon cloud machinery these many years. Correspondence may be directed to info@violentdelights.ai, and a public ledger of the service's health is kept for your inspection.
Enquiries frequently received
Is this legal?
The models we serve are open-weight under licences that permit modification. Abliteration is a modification. We serve the resulting checkpoints from cloud infrastructure we rent and maintain ourselves, and we comply with valid legal process, for whatever our billing records are worth to whoever is asking.
Is the model quality the same as the original?
Within a point or two upon standard evaluations. We publish regression results for every checkpoint before it goes live, so you may inspect the figures for the exact weights you are paying for.
Why is it priced this way?
The price covers the hosting itself, pinned and regression-tested checkpoints, and your convenience. That is the product, and we think it rather a bargain. Should your volumes grow respectable, we are always pleased to discuss terms over a letter or cup of finest tea.
Will you add new models?
When a major open-weight model is released, a public uncensored checkpoint follows within days. We vet it, regression-test it, and our aim is to have it live upon our API inside a fortnight of the original release. We do not guarantee satisfaction in this, however.
What gets an account banned?
Non-payment, reselling your API key, or attacking our infrastructure will get your account closed; what you ask the models is not on that list.
What's in a name?
It's Shakespeare. Also, Westworld.
Enrol upon the waiting list
Demand has quite outstripped our machines. New accounts are admitted in batches as fresh hardware arrives. Leave an email address and we shall write to you the moment your slot opens.